FOCUS: A Subspecialty Foundation Model for Intraoperative Central Nervous System Pathology
Authors: Xing Liu, Yifan Guo, Hongtai Lv, Lei Lou, Yuyi Tan, Jiang Du, Xijia Zhang, Ming Zhang, +13 more
Organizations: Department of Neuropathology, Beijing Neurosurgical Institute, Capital Medical University, No. 119 South Fourth Ring West Road, Fengtai District, Beijing,100070, P.R.China. · CAS Key Laboratory of Molecular Imaging, Institute of Automation, Chinese Academy of Sciences, No. 95 Zhongguancun East Road, Haidian District, Beijing, 100190, P.R. China. · School of Artificial Intelligence, University of Chinese Academy of Sciences, No. 1 Yanqihu East Road, Huairou District, Beijing, 101408, P.R. China. · Department of Pathology, The First Hospital of Hebei Medical University, No. 89 Donggang Road, Shijiazhuang City, Hebei Province, 050031, P.R. China. · Department of Neurosurgery, The First Hospital of Hebei Medical University, No. 89 Donggang Road, Shijiazhuang City, Hebei Province, 050031, P.R. China. · School of Advanced Interdisciplinary Sciences, University of Chinese Academy of Sciences, No. 1 Yanqihu East Road, Huairou District, Beijing, 101408, P.R. China. · Guangdong Provincial People’s Hospital, No. 106 Zhongshan 2nd Road, Guangzhou, Guangdong Province, 510180, P.R. China. · Beijing Advanced Innovation Center for Big Data-Based Precision Medicine, School of Engineering Medicine, Beihang University, No. 37 Xueyuan Road, Haidian District, Beijing 100191, P.R. China.
Intraoperative frozen-section diagnosis of central nervous system (CNS) lesions guides surgical decisions but is challenged by artefacts, diagnostic complexity and time constraints. Most pathology foundation models have limited exposure to CNS frozen tissue. Here we present FOCUS, a subspecialty foundation model pretrained on 27,930 frozen-sections enriched for CNS tissue. FOCUS led 11 internal diagnostic and 12 biomarker-prediction tasks, with mean AUCs of 0.969 and 0.877, respectively, retaining leading performance in temporal and external cohorts. To provide more comprehensive diagnostic information, we further organized ten diagnostic tasks into a WHO-aligned diagnostic hierarchy and separately evaluated FOCUS for lymphoma detection, a critical task for intraoperative surgical decision-making. In reader assessments by six neuropathologists, FOCUS assistance enabled junior readers to approach or exceed unassisted senior readers in hierarchical diagnosis and raised junior readers' lymphoma sensitivity from 81.6% to 94.7%. FOCUS could improve diagnostic accuracy and granularity, inform lymphoma-related resection decisions and provide provisional molecular information. Such capabilities highlight the potential of FOCUS to provide robust, clinically actionable support for intraoperative CNS pathology.
Figures & tables
Figure 1: Overview of FOCUS. a, Study workflow from data collection and model training to diagnostic, molecular and clinical evaluation. b, Distribution of 27,930 pretraining whole-slide images (WSIs) across seven public datasets. c, Distribution across 27 anatomical categories; brain specimens accounted for 49.0% of the pretraining WSIs. d, Multi-scale patch curation, UNI encoding and hierarchical sampling from 100,000 fine-grained and 10,000 top-level clusters, yielding approximately 11 million patches. e, Teacher–student adaptation using low-rank adaptation (LoRA) and Frozen-Aware Multi-View (FAMV) augmentation. f, AUC across 11 diagnostic tasks (left) and 12 biomarker-prediction tasks (right) for FOCUS and eight pathology foundation models in the internal test cohort; boxed values indicate FOCUS performance. g, Number of diagnostic and biomarker-prediction tasks on which each model ranked first in the independent temporal validation cohort. h, Number of diagnostic and biomarker-prediction tasks in which each model ranked first in the external validation cohort. For tasks in which AUC could not be calculated because one class was absent, accuracy was used instead.
Figure 2: Diagnostic performance of FOCUS for intraoperative CNS pathology. a, AUC for binary tasks or macro-AUC for multiclass tasks across 11 diagnostic tasks in the internal cohort; error bars denote 95% confidence intervals and horizontal bars show class counts. Statistical comparisons used two-sided paired Wilcoxon signed-rank tests. b, Few-shot AUCs using 1–32 examples per class across five repeats. Boxes show interquartile ranges, centre lines show medians, and whiskers extend to 1.5 times the interquartile range. c,d, Mean ROC curves across the 11 tasks in the temporal and external cohorts, respectively. Comparisons between FOCUS and each baseline used two-sided paired Wilcoxon signed-rank tests across the 11 tasks. e, AUC for lymphoma differentiation in the dedicated lymphoma differentiation cohort. f, Accuracy for lymphoma differentiation in the dedicated lymphoma differentiation cohort. g, Specificity at 95% sensitivity for tumour detection across the internal test, temporal validation and external validation cohorts. h, Specificity at 95% sensitivity after pooling slide-level predictions from the internal test, temporal validation, external validation and lymphoma differentiation cohorts. i,j, Attention maps and representative high- and low-attention regions for brain tissue versus tumour and lymphoma differentiation. Red and blue outlines indicate high- and low-attention regions, respectively. All statistical tests were corrected for multiple comparisons using the Benjamini–Hochberg procedure. *Adjusted P<0.05 , **adjusted P<0.01 and ***adjusted P<0.001 .
Figure 3: Hierarchical diagnostic performance of FOCUS. a, Ten of the 11 diagnostic tasks were organized into a four-level central nervous system (CNS) tumour hierarchy. In the internal test set, heatmaps show cumulative level-wise leaf accuracy (ACC) and hierarchical F1 score (H-F1) after truncation at each level; lower bar plots show mean maximum correct depth and depth-weighted hierarchical F1 score (DH-F1). b,c, H-F1, DH-F1 and level-wise ACC in the temporal ( n=309 slides) and independent external ( n=246 slides) validation cohorts, respectively. d,e, H-F1 and DH-F1 in diagnostically uncertain ( n=77 cases) and diagnostically discordant ( n=66 cases) subsets, respectively. f, Changes in task-level macro-AUC after adding age and tumour location to morphology. g, Changes in hierarchical metrics and level-wise ACC after clinical-information integration. Points and error bars in b–e indicate point estimates and 95% confidence intervals.
Figure 4: FOCUS-assisted hierarchical diagnosis in the reader study. Six neuropathologists (three junior and three senior) independently reviewed 243 cases and then completed an AI-assisted review after a two-week washout period. The reader-study cohort comprised 100 randomly selected cases, 77 cases whose intraoperative reports contained explicit diagnostic uncertainty and 66 cases with discordance between intraoperative and final diagnoses. a, H-F1 for FOCUS, unaided readers and AI-assisted readers; dashed lines denote group means. b, Mean DH-F1 across the three case cohorts and reader groups. c, Mean maximum correct classification depth for FOCUS and unaided readers. d, Reader-specific change in maximum correct depth after AI assistance; points and error bars show mean changes and 95% confidence intervals. e, Mean level-wise leaf accuracy for unaided and AI-assisted junior and senior readers. f, Case-level diagnostic changes after AI assistance. Blue indicates wrong-to-correct conversions, purple indicates deeper correct classifications, and light blue indicates correct-to-wrong changes. g,h, Changes in reader accuracy and recall, respectively, before and after AI assistance in the lymphoma differentiation cohort. Circles and diamonds denote performance without and with AI assistance, respectively; green labels indicate absolute percentage-point changes, and dashed horizontal lines show the performance of AI alone.
Figure 5: Biomarker-prediction performance and prototype morphology of FOCUS. a, AUCs for 12 biomarker-prediction tasks in the internal test cohort. Error bars denote 95% confidence intervals; horizontal bars show class counts. Statistical comparisons used two-sided paired Wilcoxon signed-rank tests. b, Few-shot AUCs using 1–32 examples per class across five repeats; validation and test sets were fixed. Boxes show interquartile ranges, centre lines show medians, and whiskers extend to 1.5 times the interquartile range. c,d, Mean ROC curves in the temporal and independent external cohorts, respectively. Comparisons between FOCUS and each baseline used two-sided paired Wilcoxon signed-rank tests across the evaluable tasks. e,f, Representative patches from ten shared prototypes and their UMAP projection. g, Mean slide-level prototype-proportion differences for IDH mutation, 1p/19q co-deletion within IDH-mutant gliomas, TERT mutation within IDH-wild-type gliomas and CDKN2A/B deletion. Positive values indicate enrichment in the molecular-positive group. Prototype proportions were compared using two-sided Mann–Whitney U tests. All statistical tests were corrected for multiple comparisons using the Benjamini–Hochberg procedure. *Adjusted P<0.05 , **adjusted P<0.01 and ***adjusted P<0.001 .
Figure 6: Transferability and robustness of FOCUS representations. a, AUC differences between FOCUS and seven pathology foundation models across pan-cancer tasks with external (left) or internal-only (right) validation. Differences are calculated as FOCUS minus comparator, such that positive values favour FOCUS. b,c, Win–tie–loss summaries for externally and internally validated tasks, respectively; ties denote ∣ΔAUC∣≤10−6 . d, Five-fold linear-probe accuracy (mean ± s.d.) for six Ivy-GAP histological regions sampled from 20 pretraining-excluded WSIs; all encoders were frozen. e, Paired AUC changes across 23 tasks for 448- and 896-pixel inputs relative to the 224-pixel reference; coloured markers show mean changes with 95% task-bootstrap confidence intervals. P values are from two-sided paired Wilcoxon signed-rank tests across the 23 tasks. f, Paired changes in test AUC attributable to hierarchical clustering-based data curation (HCDC) and FAMV, calculated against the corresponding ablation condition without each component; dots denote tasks and diamonds with bars denote means with 95% task-bootstrap confidence intervals. P values are from two-sided paired Wilcoxon signed-rank tests across the 23 matched tasks and are reported without multiplicity correction.
Appendix figures & tables95 assets
Supplementary material from the paper’s appendix.
Appendix
Extended Data Figure 1: Patient selection and cohort composition. Flow diagrams summarize the inclusion and exclusion of patients who underwent surgical treatment for central nervous system (CNS) tumours.
Extended Data Figure 2: Few-shot performance across diagnostic classification tasks. Box plots compare the area under the receiver operating characteristic curve (AUC) of nine pathology foundation models at 1, 2, 4, 8, 16 and 32-shot settings. Panels correspond to a, Non-neoplastic brain tissue vs tumour; b, Metastasis vs other CNS tumours; c, Neuroepithelial vs non-neuroepithelial tumour; d, Glioma or glioneuronal tumour vs embryonal tumour vs ependymoma; e, Glioma vs glioneuronal tumour; f, Meningioma vs other non-neuroepithelial tumours; g, PitNET vs SFT/mesenchymal tumour vs lymphoma vs craniopharyngioma vs schwannoma; h, Diffuse vs circumscribed glioma; i, Astrocytoma, IDH-mutant vs oligodendroglioma, IDH-mutant and 1p/19q-codeleted vs glioblastoma, IDH-wildtype vs diffuse hemispheric glioma, H3 G34-mutant vs diffuse midline glioma, H3 K27-altered; j, Pilocytic astrocytoma vs pleomorphic xanthoastrocytoma; k, Lymphoma differentiation. Boxes indicate the interquartile range, centre lines denote medians, whiskers extend to 1.5 times the interquartile range, and dots represent individual sampling runs. The shared AUC axis ranges from 0.25 to 1.00, and models are identified by the shared colour legend.
Extended Data Figure 3: Few-shot performance across biomarker prediction tasks. Box plots compare the area under the receiver operating characteristic curve (AUC) of nine pathology foundation models at 1, 2, 4, 8, 16 and 32-shot settings. Panels correspond to biomarker prediction tasks for a, ATRX alteration; b, 1p/19q codeletion; c, H3 K27 alteration; d, H3 G34 mutation; e, BRAF mutation; f, CDKN2A/B deletion; g, TP53 mutation; h, IDH1/IDH2 mutation; i, MGMT promoter methylation; j, chromosome 7 gain/chromosome 10 loss (+7/-10); k, EGFR amplification; and l, TERT promoter mutation. Boxes indicate the interquartile range, centre lines denote medians, whiskers extend to 1.5 times the interquartile range, and dots represent individual sampling runs. The shared AUC axis ranges from 0.25 to 1.00, and models are identified by the shared colour legend.
Extended Data Figure 4: Representative attention-map examples for diagnostic classification tasks. Each subpanel shows a WSI thumbnail, the corresponding attention heatmap, two high-response patches (red frames), and two low-response patches (blue frames). Panel-task mapping: a, neuroepithelial versus non-neuroepithelial tumour classification; b, glioma or glioneuronal tumour versus embryonal tumour versus ependymoma classification; c, glioma versus glioneuronal tumour classification; d, meningioma versus other non-neuroepithelial tumours classification; e, PitNET versus SFT/mesenchymal tumour versus lymphoma versus craniopharyngioma versus schwannoma classification; f, diffuse versus circumscribed glioma classification; g, diffuse glioma molecular-subtype classification; h, pilocytic astrocytoma versus pleomorphic xanthoastrocytoma classification; i, lymphoma differentiation.
Extended Data Figure 5: Representative attention-map examples for biomarker prediction tasks. Each subpanel shows a WSI thumbnail, the corresponding attention heatmap, two high-response patches (red frames), and two low-response patches (blue frames). Panel-task mapping: a, ATRX alteration prediction; b, 1p/19q codeletion prediction; c, H3 K27 alteration prediction; d, H3 G34 mutation prediction; e, BRAF mutation prediction; f, CDKN2A/B deletion prediction; g, TP53 mutation prediction; h, IDH1/IDH2 mutation prediction; i, MGMT promoter methylation prediction; j, chromosome 7 gain/chromosome 10 loss (+7/-10) prediction; k, EGFR amplification prediction; l, TERT promoter mutation prediction.
Extended Data Figure 6: Representative correct and incorrect AI predictions in lymphoma differentiation. Panels show ADC and T1-FLAIR MRI, frozen sections, AI attention maps, and preoperative, intraoperative, postoperative and AI diagnoses. a, A right cerebellar lesion favoured glioma on imaging, while frozen pathology could not exclude high-grade glioma or embryonal tumour. Final pathology confirmed germinal-centre DLBCL, correctly identified as lymphoma by AI. b, A parietal lesion suggested metastasis on imaging and high-grade glioma on frozen section. Final pathology confirmed non-germinal-centre DLBCL, correctly identified as lymphoma by AI. c, Imaging suggested glioma and frozen pathology lymphoma or embryonal tumour; final pathology confirmed CNS WHO grade 4 glioma, misclassified as lymphoma by AI. Green and red markings indicate concordance and discordance with final pathology, respectively.
Extended Data Figure 7: Prototype enrichment for molecular alterations. Prototype-proportion differences are shown for ATRX alteration in IDH-mutant gliomas, H3 K27 alteration, H3 G34 mutation, BRAF mutation in IDH-wild-type gliomas, chromosome 7 gain/chromosome 10 loss in IDH-wild-type gliomas, and EGFR amplification in IDH-wild-type gliomas. Bars show the mean prototype-proportion difference between molecular-positive and molecular-negative slides. Salmon and blue bars indicate positive and negative enrichment, respectively. Each panel uses the shared ten-prototype dictionary; annotations report the corresponding enrichment-test P values.
Data source
Slides
Percentage (%)
TCGA
13,807
49.43
Ivy-GAP
11,230
40.21
HEST
1,228
4.40
CGCI
763
2.73
HCMI-CMDC
439
1.57
CPTAC
414
1.48
Appendix
Extended Data Table 1: Distribution of pretraining slides by data source.
Organ
Slides
Percentage (%)
Brain
13,689
49.01
Kidney
1,725
6.18
Bronchus and lungs
1,672
5.99
Breast
1,525
5.46
Bowel
1,159
4.15
Ovary
1,048
3.75
Appendix
Extended Data Table 2: Distribution of pretraining slides by organ.
Tasks
Training cohort
Test cohort
Temporal cohort
External cohort
Non-neoplastic brain tissue/tumour
2777 (200/2577)
694 (54/640)
309 (29/280)
246 (9/237)
Metastasis/ other CNS tumours
2577 (112/2465)
640 (28/612)
280 (14/266)
237 (13/224)
Neuroepithelial/ non-neuroepithelial tumour
2291 (1876/415)
576 (472/104)
255 (202/53)
221 (58/163)
Glioma or GNT/embryonal tumour/ ependymoma
1876 (1700/90/86)
472 (427/25/20)
202 (176/17/9)
58 (52/0/6)
Glioma/GNT
1700 (1553/147)
427 (390/37)
176 (162/14)
52 (50/2)
Meningioma/ other non-neuroepithelial tumours
415 (246/169)
104 (62/42)
53 (35/18)
163 (72/91)
Appendix
Extended Data Table 3: Downstream task partitions and label distributions in the development cohort and two independent validation cohorts.
Data source
Slides
Percentage (%)
TCGA
6,336
53.89
CPTAC
5,049
42.94
HCMI-CMDC
165
1.40
CGCI
208
1.77
Total
11758
100.00
Appendix
Extended Data Table 4: Distribution of downstream tasks slides by data source.
Task
Samples
Label distribution
External validation
BLCA (Benign vs. malignant)
469
37 / 432
–
BRCA (Histological subtype)
1417
1173 / 244
111 (100 / 11)
BRCA subset (Histological subtype)
408
331 / 77
115 (100 / 11 / 4)
BRCA (Benign vs. malignant)
1976
398 / 1578
–
CESC (tumour grading)
283
19 / 139 / 125
208 (10 / 102 / 96)
COADREAD (Benign vs. malignant)
1347
134 / 1213
165 (33 / 132)
Appendix
Extended Data Table 5: Summary of TCGA downstream tasks and label distributions.
Hyper-parameter
Value
Backbone architecture
ViT-G/14
Levels
40
Heads
24
Patch size
14
FFN layer
SwiGLU
Embedding dimension
1536
Appendix
Extended Data Table 6: Detailed configuration for self-supervised pretraining.
Task
Definition
T1
Non-neoplastic brain tissue versus tumour classification
T2
Metastasis versus other CNS tumours classification
T3
Neuroepithelial versus non-neuroepithelial tumour classification
T4
Glioma or glioneuronal tumour versus embryonal tumour versus ependymoma classification
T5
Glioma versus glioneuronal tumour classification
T6
Meningioma versus other non-neuroepithelial tumours classification
Appendix
Extended Data Table 7: Definitions of the 23 diagnostic and molecular classification tasks.
Task
DSMIL
ABMIL
CLAM
TransMIL
T1
0.991 (0.984–0.998)
0.990 (0.982–0.998)
0.991 (0.983–0.998)
0.989 (0.981–0.997)
T2
0.992 (0.986–0.999)
0.976 (0.954–0.997)
0.990 (0.981–0.998)
0.994 (0.989–0.999)
T3
0.999 (0.998–1.000)
0.996 (0.991–1.000)
0.997 (0.993–1.000)
0.999 (0.998–1.000)
T4
0.974 (0.929–0.997)
0.975 (0.941–0.995)
0.979 (0.948–0.996)
0.985 (0.960–0.999)
T5
0.946 (0.915–0.976)
0.947 (0.918–0.977)
0.952 (0.923–0.981)
0.946 (0.914–0.978)
T6
0.998 (0.994–1.000)
0.995 (0.984–1.000)
0.994 (0.982–1.000)
0.996 (0.989–1.000)
Appendix
Extended Data Table 8: Internal validation set AUC and 95% confidence intervals for FOCUS across MIL methods.
Task
DSMIL
ABMIL
CLAM
TransMIL
T1
0.964 (0.943–0.985)
0.974 (0.960–0.989)
0.970 (0.954–0.986)
0.968 (0.951–0.986)
T2
0.995 (0.991–1.000)
0.983 (0.964–1.000)
0.993 (0.985–1.000)
0.995 (0.989–1.000)
T3
0.992 (0.986–0.998)
0.991 (0.984–0.999)
0.992 (0.984–0.999)
0.988 (0.979–0.997)
T4
0.987 (0.977–0.995)
0.985 (0.971–0.995)
0.985 (0.973–0.995)
0.985 (0.971–0.995)
T5
0.898 (0.861–0.936)
0.902 (0.865–0.939)
0.917 (0.883–0.950)
0.908 (0.869–0.947)
T6
0.992 (0.979–1.000)
0.988 (0.972–1.000)
0.987 (0.970–1.000)
0.987 (0.973–1.000)
Appendix
Extended Data Table 9: Internal test set AUC and 95% confidence intervals for FOCUS across MIL methods.
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.964 (0.943–0.985)
0.947 (0.928–0.963)
0.963 (0.946–0.976)
0.759 (0.635–0.878)
0.971 (0.960–0.980)
H-optimus-1
0.959 (0.930–0.988)
0.935 (0.915–0.954)
0.941 (0.921–0.959)
0.870 (0.771–0.957)
0.964 (0.952–0.974)
H-optimus-0
0.954 (0.923–0.986)
0.937 (0.916–0.954)
0.942 (0.922–0.960)
0.870 (0.771–0.957)
0.965 (0.953–0.975)
GigaPath
0.958 (0.931–0.984)
0.935 (0.915–0.952)
0.944 (0.923–0.961)
0.833 (0.727–0.929)
0.964 (0.952–0.974)
CRISP
0.957 (0.929–0.985)
0.935 (0.915–0.952)
0.947 (0.927–0.964)
0.796 (0.678–0.900)
0.964 (0.953–0.974)
Virchow2
0.958 (0.927–0.989)
0.954 (0.937–0.968)
0.969 (0.954–0.981)
0.778 (0.653–0.889)
0.975 (0.965–0.983)
Appendix
Extended Data Table 10: Performance of various models on internal test set for T1: Normal brain (0) vs tumour (1).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.995 (0.991–1.000)
0.984 (0.975–0.994)
0.997 (0.992–1.000)
0.714 (0.520–0.879)
0.992 (0.987–0.997)
H-optimus-1
0.992 (0.985–0.999)
0.983 (0.972–0.992)
0.995 (0.989–1.000)
0.714 (0.538–0.875)
0.991 (0.985–0.996)
H-optimus-0
0.989 (0.980–0.998)
0.981 (0.970–0.991)
0.995 (0.989–1.000)
0.679 (0.500–0.850)
0.990 (0.984–0.995)
GigaPath
0.981 (0.968–0.994)
0.973 (0.961–0.984)
0.995 (0.989–1.000)
0.500 (0.308–0.690)
0.986 (0.980–0.992)
CRISP
0.986 (0.973–0.998)
0.980 (0.969–0.989)
0.998 (0.995–1.000)
0.571 (0.385–0.750)
0.989 (0.984–0.994)
Virchow2
0.980 (0.963–0.997)
0.983 (0.972–0.992)
0.998 (0.995–1.000)
0.643 (0.458–0.815)
0.991 (0.985–0.996)
Appendix
Extended Data Table 11: Performance of various models on internal test set for T2: Metastasis (0) vs all other tumours (1).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.992 (0.986–0.998)
0.973 (0.959–0.988)
0.923 (0.867–0.971)
0.985 (0.972–0.996)
0.928 (0.888–0.964)
H-optimus-1
0.989 (0.981–0.997)
0.973 (0.959–0.986)
0.933 (0.880–0.979)
0.983 (0.969–0.993)
0.928 (0.889–0.963)
H-optimus-0
0.985 (0.974–0.996)
0.968 (0.952–0.982)
0.913 (0.854–0.967)
0.980 (0.966–0.991)
0.913 (0.868–0.951)
GigaPath
0.985 (0.976–0.995)
0.966 (0.950–0.980)
0.923 (0.867–0.971)
0.976 (0.960–0.989)
0.910 (0.864–0.948)
CRISP
0.985 (0.973–0.996)
0.973 (0.959–0.986)
0.933 (0.880–0.980)
0.983 (0.969–0.994)
0.928 (0.889–0.962)
Virchow2
0.986 (0.973–0.998)
0.972 (0.957–0.984)
0.923 (0.867–0.972)
0.983 (0.970–0.993)
0.923 (0.882–0.958)
Appendix
Extended Data Table 12: Performance of various models on internal test set for T3: Neuroepithelial (0) vs non-neuroepithelial tumours (1).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.987 (0.977–0.995)
0.974 (0.959–0.986)
0.972 (0.958–0.985)
0.779 (0.668–0.887)
0.971 (0.953–0.985)
H-optimus-1
0.983 (0.972–0.993)
0.958 (0.938–0.974)
0.956 (0.934–0.972)
0.638 (0.502–0.764)
0.951 (0.926–0.971)
H-optimus-0
0.977 (0.958–0.991)
0.964 (0.946–0.980)
0.962 (0.943–0.979)
0.678 (0.550–0.803)
0.957 (0.933–0.976)
GigaPath
0.977 (0.959–0.991)
0.966 (0.950–0.980)
0.964 (0.947–0.979)
0.718 (0.590–0.835)
0.961 (0.940–0.979)
CRISP
0.974 (0.945–0.992)
0.960 (0.942–0.976)
0.958 (0.939–0.975)
0.658 (0.529–0.782)
0.954 (0.930–0.973)
Virchow2
0.977 (0.955–0.992)
0.962 (0.944–0.978)
0.960 (0.941–0.977)
0.658 (0.535–0.783)
0.953 (0.929–0.974)
Appendix
Extended Data Table 13: Performance of various models on internal test set for T4: Glioma or glioneuronal tumour (0) vs embryonal tumour (1) vs ependymoma (2).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.898 (0.861–0.936)
0.899 (0.871–0.925)
0.378 (0.216–0.533)
0.949 (0.926–0.969)
0.394 (0.235–0.531)
H-optimus-1
0.869 (0.822–0.917)
0.909 (0.881–0.934)
0.216 (0.094–0.359)
0.974 (0.957–0.990)
0.291 (0.130–0.453)
H-optimus-0
0.855 (0.804–0.906)
0.902 (0.874–0.927)
0.162 (0.056–0.286)
0.972 (0.954–0.987)
0.222 (0.078–0.370)
GigaPath
0.863 (0.816–0.911)
0.888 (0.857–0.916)
0.135 (0.029–0.256)
0.959 (0.938–0.977)
0.172 (0.038–0.306)
CRISP
0.887 (0.845–0.928)
0.892 (0.862–0.918)
0.514 (0.345–0.677)
0.928 (0.901–0.951)
0.452 (0.311–0.575)
Virchow2
0.880 (0.836–0.924)
0.892 (0.859–0.920)
0.459 (0.294–0.621)
0.933 (0.907–0.957)
0.425 (0.276–0.556)
Appendix
Extended Data Table 14: Performance of various models on internal test set for T5: Glioma (0) vs glioneuronal tumour (1).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.992 (0.979–1.000)
0.971 (0.933–1.000)
0.976 (0.921–1.000)
0.968 (0.919–1.000)
0.965 (0.919–1.000)
H-optimus-1
0.982 (0.961–1.000)
0.952 (0.904–0.990)
0.929 (0.844–1.000)
0.968 (0.919–1.000)
0.940 (0.880–0.987)
H-optimus-0
0.974 (0.947–1.000)
0.923 (0.865–0.971)
0.905 (0.809–0.978)
0.935 (0.866–0.985)
0.905 (0.831–0.961)
GigaPath
0.984 (0.967–1.000)
0.952 (0.904–0.990)
0.952 (0.879–1.000)
0.952 (0.889–1.000)
0.941 (0.878–0.987)
CRISP
0.988 (0.975–1.000)
0.923 (0.865–0.971)
0.905 (0.809–0.978)
0.935 (0.869–0.985)
0.905 (0.831–0.963)
Virchow2
0.986 (0.971–1.000)
0.942 (0.894–0.981)
0.905 (0.806–0.977)
0.968 (0.919–1.000)
0.927 (0.857–0.978)
Appendix
Extended Data Table 15: Performance of various models on internal test set for T6: Meningioma (0) vs other non-neuroepithelial tumours (1).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.977 (0.944–0.997)
0.925 (0.867–0.972)
0.833 (0.714–0.929)
0.938 (0.884–0.979)
0.830 (0.707–0.929)
H-optimus-1
0.971 (0.937–0.993)
0.900 (0.834–0.954)
0.786 (0.667–0.905)
0.914 (0.853–0.962)
0.775 (0.627–0.899)
H-optimus-0
0.958 (0.917–0.986)
0.881 (0.815–0.935)
0.738 (0.595–0.857)
0.899 (0.833–0.948)
0.729 (0.573–0.856)
GigaPath
0.968 (0.927–0.994)
0.921 (0.859–0.969)
0.810 (0.667–0.929)
0.939 (0.885–0.980)
0.803 (0.655–0.923)
CRISP
0.976 (0.944–0.997)
0.927 (0.869–0.972)
0.833 (0.714–0.929)
0.956 (0.917–0.987)
0.836 (0.723–0.931)
Virchow2
0.971 (0.930–0.995)
0.894 (0.822–0.950)
0.762 (0.619–0.881)
0.933 (0.889–0.971)
0.762 (0.625–0.881)
Appendix
Extended Data Table 16: Performance of various models on internal test set for T7: PitNET (0) vs SFT/mesenchymal tumour (1) vs lymphoma (2) vs craniopharyngioma (3) vs schwannoma (4).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.956 (0.933–0.979)
0.921 (0.893–0.948)
0.600 (0.473–0.733)
0.976 (0.957–0.991)
0.688 (0.575–0.787)
H-optimus-1
0.945 (0.916–0.974)
0.906 (0.874–0.935)
0.491 (0.360–0.623)
0.976 (0.959–0.991)
0.600 (0.474–0.713)
H-optimus-0
0.939 (0.905–0.972)
0.914 (0.882–0.942)
0.527 (0.393–0.661)
0.979 (0.963–0.994)
0.637 (0.511–0.741)
GigaPath
0.948 (0.919–0.977)
0.914 (0.885–0.942)
0.600 (0.467–0.731)
0.966 (0.945–0.985)
0.667 (0.548–0.766)
CRISP
0.939 (0.910–0.968)
0.908 (0.880–0.937)
0.655 (0.524–0.778)
0.951 (0.928–0.973)
0.673 (0.560–0.768)
Virchow2
0.935 (0.902–0.967)
0.908 (0.880–0.935)
0.618 (0.491–0.750)
0.957 (0.934–0.978)
0.660 (0.548–0.759)
Appendix
Extended Data Table 17: Performance of various models on internal test set for T8: Diffuse (0) vs circumscribed glioma (1).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.932 (0.912–0.950)
0.869 (0.843–0.893)
0.761 (0.716–0.804)
0.884 (0.860–0.906)
0.750 (0.699–0.796)
H-optimus-1
0.911 (0.889–0.932)
0.856 (0.830–0.881)
0.734 (0.688–0.780)
0.876 (0.855–0.897)
0.716 (0.663–0.765)
H-optimus-0
0.893 (0.868–0.918)
0.852 (0.825–0.876)
0.728 (0.679–0.774)
0.869 (0.846–0.891)
0.709 (0.655–0.760)
GigaPath
0.916 (0.894–0.936)
0.861 (0.833–0.885)
0.749 (0.700–0.793)
0.876 (0.851–0.899)
0.737 (0.683–0.784)
CRISP
0.908 (0.884–0.930)
0.845 (0.817–0.870)
0.719 (0.670–0.765)
0.861 (0.836–0.884)
0.696 (0.641–0.746)
Virchow2
0.905 (0.880–0.928)
0.850 (0.822–0.876)
0.731 (0.682–0.777)
0.865 (0.838–0.889)
0.715 (0.660–0.766)
Appendix
Extended Data Table 18: Performance of various models on internal test set for T9: IDH-mutant astrocytoma (0), IDH-mutant and 1p/19q-codeleted oligodendroglioma (1), IDH-wild-type GBM (2), H3G34-mutant DHG (3), or H3K27-altered DMG (4).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.988 (0.965–1.000)
0.930 (0.860–0.982)
0.923 (0.750–1.000)
0.932 (0.848–1.000)
0.857 (0.667–0.970)
H-optimus-1
0.963 (0.918–1.000)
0.930 (0.860–0.982)
0.769 (0.500–1.000)
0.977 (0.927–1.000)
0.833 (0.625–0.968)
H-optimus-0
0.930 (0.862–0.998)
0.877 (0.789–0.965)
0.692 (0.400–0.923)
0.932 (0.850–1.000)
0.720 (0.454–0.897)
GigaPath
0.965 (0.921–1.000)
0.930 (0.860–0.982)
0.846 (0.611–1.000)
0.955 (0.881–1.000)
0.846 (0.640–0.970)
CRISP
0.962 (0.914–1.000)
0.895 (0.807–0.965)
0.769 (0.500–1.000)
0.932 (0.848–1.000)
0.769 (0.533–0.923)
Virchow2
0.932 (0.857–1.000)
0.895 (0.807–0.965)
0.692 (0.400–0.929)
0.955 (0.886–1.000)
0.750 (0.500–0.917)
Appendix
Extended Data Table 19: Performance of various models on internal test set for T10: Pilocytic astrocytoma (0) vs pleomorphic xanthoastrocytoma (1).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.973 (0.937–1.000)
0.987 (0.976–0.996)
0.400 (0.100–0.714)
1.000 (1.000–1.000)
0.571 (0.182–0.833)
H-optimus-1
0.963 (0.913–1.000)
0.989 (0.978–0.998)
0.500 (0.167–0.833)
1.000 (1.000–1.000)
0.667 (0.286–0.909)
H-optimus-0
0.929 (0.816–1.000)
0.989 (0.978–0.998)
0.500 (0.167–0.833)
1.000 (1.000–1.000)
0.667 (0.286–0.909)
GigaPath
0.952 (0.888–1.000)
0.989 (0.978–0.998)
0.500 (0.167–0.833)
1.000 (1.000–1.000)
0.667 (0.286–0.909)
CRISP
0.956 (0.891–1.000)
0.993 (0.985–1.000)
0.700 (0.400–1.000)
1.000 (1.000–1.000)
0.824 (0.571–1.000)
Virchow2
0.924 (0.809–1.000)
0.985 (0.971–0.993)
0.500 (0.182–0.833)
0.995 (0.989–1.000)
0.588 (0.250–0.833)
Appendix
Extended Data Table 20: Performance of various models on internal test set for T11: Glioma, metastasis or embryonal tumour (0) vs lymphoma (1).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.834 (0.779–0.890)
0.831 (0.791–0.868)
0.592 (0.474–0.697)
0.897 (0.860–0.933)
0.604 (0.500–0.688)
H-optimus-1
0.819 (0.760–0.878)
0.831 (0.791–0.868)
0.539 (0.427–0.653)
0.912 (0.877–0.943)
0.582 (0.482–0.675)
H-optimus-0
0.811 (0.750–0.872)
0.834 (0.794–0.871)
0.539 (0.427–0.654)
0.916 (0.881–0.947)
0.586 (0.481–0.676)
GigaPath
0.824 (0.770–0.879)
0.828 (0.785–0.865)
0.513 (0.395–0.627)
0.916 (0.880–0.946)
0.565 (0.453–0.658)
CRISP
0.817 (0.761–0.874)
0.825 (0.785–0.863)
0.500 (0.387–0.611)
0.916 (0.880–0.947)
0.555 (0.448–0.650)
Virchow2
0.808 (0.751–0.866)
0.822 (0.782–0.862)
0.421 (0.309–0.535)
0.934 (0.903–0.961)
0.508 (0.393–0.612)
Appendix
Extended Data Table 21: Performance of various models on internal test set for T12: ATRX unaltered (0) vs altered (1).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.967 (0.951–0.984)
0.911 (0.883–0.940)
0.720 (0.625–0.813)
0.970 (0.948–0.989)
0.792 (0.719–0.859)
H-optimus-1
0.954 (0.931–0.978)
0.909 (0.877–0.937)
0.720 (0.618–0.815)
0.966 (0.944–0.985)
0.787 (0.708–0.855)
H-optimus-0
0.946 (0.921–0.971)
0.889 (0.854–0.920)
0.610 (0.500–0.716)
0.974 (0.954–0.992)
0.719 (0.628–0.800)
GigaPath
0.951 (0.929–0.974)
0.897 (0.866–0.929)
0.683 (0.581–0.784)
0.963 (0.939–0.982)
0.757 (0.672–0.827)
CRISP
0.933 (0.905–0.962)
0.891 (0.860–0.923)
0.622 (0.518–0.727)
0.974 (0.954–0.992)
0.729 (0.642–0.808)
Virchow2
0.924 (0.891–0.957)
0.877 (0.843–0.911)
0.610 (0.506–0.720)
0.959 (0.933–0.981)
0.699 (0.611–0.783)
Appendix
Extended Data Table 22: Performance of various models on internal test set for T13: 1p/19q non-codeleted (0) vs codeleted (1).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.905 (0.858–0.951)
0.945 (0.919–0.968)
0.176 (0.000–0.375)
0.985 (0.970–0.997)
0.240 (0.000–0.455)
H-optimus-1
0.815 (0.722–0.908)
0.945 (0.919–0.968)
0.059 (0.000–0.188)
0.991 (0.979–1.000)
0.095 (0.000–0.273)
H-optimus-0
0.816 (0.719–0.913)
0.942 (0.916–0.965)
0.059 (0.000–0.188)
0.988 (0.975–0.997)
0.091 (0.000–0.267)
GigaPath
0.833 (0.750–0.917)
0.948 (0.922–0.968)
0.176 (0.000–0.375)
0.988 (0.975–0.997)
0.250 (0.000–0.471)
CRISP
0.791 (0.697–0.884)
0.957 (0.934–0.977)
0.294 (0.083–0.533)
0.991 (0.979–1.000)
0.400 (0.125–0.621)
Virchow2
0.753 (0.638–0.867)
0.945 (0.919–0.968)
0.176 (0.000–0.375)
0.985 (0.970–0.997)
0.240 (0.000–0.457)
Appendix
Extended Data Table 23: Performance of various models on internal test set for T14: H3K27 unaltered (0) vs altered (1).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.981 (0.962–1.000)
0.986 (0.971–0.997)
0.250 (0.000–1.000)
0.994 (0.986–1.000)
0.286 (0.000–0.667)
H-optimus-1
0.974 (0.941–1.000)
0.986 (0.971–0.994)
0.000 (0.000–0.000)
0.997 (0.991–1.000)
0.000 (0.000–0.000)
H-optimus-0
0.966 (0.918–1.000)
0.983 (0.969–0.994)
0.000 (0.000–0.000)
0.994 (0.986–1.000)
0.000 (0.000–0.000)
GigaPath
0.965 (0.907–1.000)
0.989 (0.977–0.997)
0.000 (0.000–0.000)
1.000 (1.000–1.000)
0.000 (0.000–0.000)
CRISP
0.981 (0.965–0.998)
0.989 (0.977–0.997)
0.000 (0.000–0.000)
1.000 (1.000–1.000)
0.000 (0.000–0.000)
Virchow2
0.978 (0.953–1.000)
0.989 (0.977–0.997)
0.000 (0.000–0.000)
1.000 (1.000–1.000)
0.000 (0.000–0.000)
Appendix
Extended Data Table 24: Performance of various models on internal test set for T15: H3G34 wild type (0) vs mutant (1).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.836 (0.722–0.950)
0.946 (0.920–0.968)
0.133 (0.000–0.333)
0.982 (0.967–0.994)
0.174 (0.000–0.385)
H-optimus-1
0.808 (0.693–0.922)
0.954 (0.931–0.974)
0.067 (0.000–0.222)
0.994 (0.985–1.000)
0.111 (0.000–0.333)
H-optimus-0
0.786 (0.656–0.916)
0.957 (0.934–0.977)
0.000 (0.000–0.000)
1.000 (1.000–1.000)
0.000 (0.000–0.000)
GigaPath
0.825 (0.706–0.943)
0.960 (0.937–0.980)
0.133 (0.000–0.333)
0.997 (0.991–1.000)
0.222 (0.000–0.467)
CRISP
0.800 (0.672–0.928)
0.954 (0.931–0.974)
0.000 (0.000–0.000)
0.997 (0.991–1.000)
0.000 (0.000–0.000)
Virchow2
0.790 (0.665–0.916)
0.957 (0.934–0.977)
0.000 (0.000–0.000)
1.000 (1.000–1.000)
0.000 (0.000–0.000)
Appendix
Extended Data Table 25: Performance of various models on internal test set for T16: BRAF wild type (0) vs mutant (1).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.823 (0.774–0.872)
0.814 (0.774–0.854)
0.517 (0.417–0.624)
0.916 (0.881–0.950)
0.586 (0.486–0.675)
H-optimus-1
0.817 (0.767–0.867)
0.797 (0.754–0.840)
0.427 (0.325–0.528)
0.923 (0.889–0.955)
0.517 (0.415–0.610)
H-optimus-0
0.808 (0.757–0.859)
0.794 (0.751–0.837)
0.438 (0.337–0.544)
0.916 (0.880–0.947)
0.520 (0.422–0.614)
GigaPath
0.802 (0.751–0.853)
0.797 (0.754–0.840)
0.427 (0.326–0.533)
0.923 (0.888–0.955)
0.517 (0.415–0.610)
CRISP
0.802 (0.750–0.853)
0.780 (0.737–0.823)
0.236 (0.152–0.326)
0.966 (0.942–0.985)
0.353 (0.245–0.457)
Virchow2
0.801 (0.749–0.854)
0.777 (0.734–0.823)
0.247 (0.159–0.337)
0.958 (0.932–0.981)
0.361 (0.244–0.466)
Appendix
Extended Data Table 26: Performance of various models on internal test set for T17: CDKN2A/B retained (0) vs deleted (1).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.866 (0.826–0.907)
0.820 (0.777–0.857)
0.737 (0.659–0.809)
0.871 (0.825–0.912)
0.757 (0.700–0.809)
H-optimus-1
0.844 (0.800–0.888)
0.803 (0.760–0.846)
0.654 (0.570–0.735)
0.894 (0.852–0.932)
0.716 (0.648–0.777)
H-optimus-0
0.837 (0.792–0.881)
0.803 (0.760–0.846)
0.654 (0.576–0.733)
0.894 (0.853–0.932)
0.716 (0.651–0.775)
GigaPath
0.843 (0.800–0.886)
0.800 (0.757–0.840)
0.662 (0.576–0.742)
0.885 (0.841–0.925)
0.715 (0.645–0.774)
CRISP
0.834 (0.790–0.878)
0.791 (0.749–0.831)
0.752 (0.675–0.820)
0.816 (0.763–0.865)
0.733 (0.672–0.786)
Virchow2
0.829 (0.783–0.875)
0.786 (0.743–0.829)
0.602 (0.520–0.682)
0.899 (0.857–0.936)
0.681 (0.611–0.745)
Appendix
Extended Data Table 27: Performance of various models on internal test set for T18: TP53 wild type (0) vs mutant (1).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.955 (0.935–0.975)
0.885 (0.848–0.919)
0.917 (0.872–0.957)
0.855 (0.800–0.907)
0.885 (0.844–0.920)
H-optimus-1
0.922 (0.892–0.952)
0.863 (0.823–0.898)
0.891 (0.839–0.936)
0.837 (0.778–0.893)
0.863 (0.818–0.900)
H-optimus-0
0.920 (0.889–0.950)
0.848 (0.807–0.888)
0.897 (0.849–0.940)
0.801 (0.737–0.861)
0.851 (0.807–0.891)
GigaPath
0.939 (0.915–0.963)
0.860 (0.820–0.898)
0.904 (0.856–0.946)
0.819 (0.757–0.877)
0.862 (0.818–0.900)
CRISP
0.927 (0.900–0.954)
0.848 (0.804–0.885)
0.865 (0.810–0.918)
0.831 (0.772–0.887)
0.846 (0.796–0.888)
Virchow2
0.927 (0.900–0.955)
0.851 (0.811–0.891)
0.865 (0.810–0.918)
0.837 (0.780–0.894)
0.849 (0.801–0.891)
Appendix
Extended Data Table 28: Performance of various models on internal test set for T19: IDH1/IDH2 wild type (0) vs either gene mutated (1).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.736 (0.677–0.794)
0.680 (0.625–0.735)
0.779 (0.714–0.843)
0.554 (0.472–0.644)
0.732 (0.673–0.782)
H-optimus-1
0.734 (0.676–0.792)
0.655 (0.600–0.709)
0.721 (0.653–0.791)
0.570 (0.483–0.658)
0.700 (0.642–0.756)
H-optimus-0
0.705 (0.644–0.767)
0.655 (0.600–0.709)
0.701 (0.631–0.773)
0.595 (0.513–0.680)
0.695 (0.632–0.748)
GigaPath
0.727 (0.668–0.787)
0.658 (0.604–0.713)
0.721 (0.652–0.789)
0.579 (0.492–0.667)
0.703 (0.642–0.757)
CRISP
0.707 (0.646–0.768)
0.644 (0.589–0.698)
0.773 (0.708–0.837)
0.479 (0.392–0.565)
0.708 (0.650–0.760)
Virchow2
0.718 (0.658–0.778)
0.673 (0.618–0.727)
0.786 (0.721–0.848)
0.529 (0.440–0.618)
0.729 (0.674–0.781)
Appendix
Extended Data Table 29: Performance of various models on internal test set for T20: MGMT unmethylated (0) vs methylated (1).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.884 (0.846–0.923)
0.838 (0.798–0.879)
0.273 (0.157–0.395)
0.955 (0.927–0.978)
0.366 (0.229–0.494)
H-optimus-1
0.836 (0.791–0.881)
0.807 (0.763–0.850)
0.127 (0.043–0.217)
0.947 (0.919–0.973)
0.184 (0.063–0.299)
H-optimus-0
0.830 (0.782–0.878)
0.810 (0.766–0.850)
0.109 (0.035–0.196)
0.955 (0.928–0.977)
0.164 (0.056–0.277)
GigaPath
0.837 (0.790–0.884)
0.816 (0.773–0.860)
0.091 (0.019–0.173)
0.966 (0.943–0.985)
0.145 (0.032–0.263)
CRISP
0.831 (0.779–0.884)
0.822 (0.782–0.863)
0.091 (0.019–0.173)
0.974 (0.953–0.992)
0.149 (0.032–0.267)
Virchow2
0.825 (0.770–0.880)
0.826 (0.782–0.866)
0.109 (0.035–0.200)
0.974 (0.954–0.989)
0.176 (0.060–0.301)
Appendix
Extended Data Table 30: Performance of various models on internal test set for T21: No concurrent chromosome 7 gain and chromosome 10 loss (0) vs concurrent gain/loss (1).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.855 (0.790–0.920)
0.888 (0.854–0.919)
0.574 (0.425–0.712)
0.942 (0.913–0.967)
0.600 (0.472–0.706)
H-optimus-1
0.828 (0.758–0.899)
0.866 (0.826–0.900)
0.489 (0.341–0.643)
0.931 (0.899–0.957)
0.517 (0.381–0.634)
H-optimus-0
0.813 (0.738–0.889)
0.866 (0.829–0.900)
0.447 (0.302–0.591)
0.938 (0.910–0.964)
0.494 (0.354–0.614)
GigaPath
0.843 (0.773–0.913)
0.872 (0.835–0.907)
0.468 (0.333–0.618)
0.942 (0.912–0.968)
0.518 (0.382–0.636)
CRISP
0.797 (0.719–0.874)
0.872 (0.835–0.907)
0.319 (0.189–0.455)
0.967 (0.944–0.986)
0.423 (0.271–0.558)
Virchow2
0.797 (0.718–0.876)
0.875 (0.838–0.910)
0.277 (0.151–0.420)
0.978 (0.960–0.993)
0.394 (0.233–0.537)
Appendix
Extended Data Table 31: Performance of various models on internal test set for T22: EGFR non-amplified (0) vs amplified (1).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.877 (0.840–0.915)
0.823 (0.780–0.866)
0.805 (0.741–0.867)
0.840 (0.782–0.897)
0.818 (0.769–0.863)
H-optimus-1
0.856 (0.816–0.897)
0.764 (0.720–0.807)
0.818 (0.756–0.875)
0.712 (0.640–0.778)
0.774 (0.724–0.821)
H-optimus-0
0.828 (0.784–0.873)
0.752 (0.705–0.798)
0.811 (0.750–0.871)
0.693 (0.622–0.765)
0.763 (0.709–0.811)
GigaPath
0.865 (0.826–0.904)
0.786 (0.739–0.829)
0.767 (0.701–0.830)
0.804 (0.742–0.863)
0.780 (0.725–0.828)
CRISP
0.852 (0.811–0.892)
0.755 (0.705–0.801)
0.774 (0.707–0.838)
0.736 (0.667–0.804)
0.757 (0.703–0.807)
Virchow2
0.829 (0.785–0.873)
0.742 (0.696–0.789)
0.799 (0.736–0.860)
0.687 (0.615–0.762)
0.754 (0.701–0.802)
Appendix
Extended Data Table 32: Performance of various models on internal test set for T23: TERT c.-124/-146 wild type (0) vs either promoter mutation (1).
Task
FOCUS
H-optimus-1
H-optimus-0
GigaPath
CRISP
Virchow2
Virchow
UNI
Phikon-v2
T1
0.954 (0.920–0.987)
0.950 (0.920–0.980)
0.949 (0.915–0.982)
0.946 (0.911–0.980)
0.942 (0.905–0.979)
0.953 (0.923–0.983)
0.941 (0.906–0.976)
0.945 (0.911–0.978)
0.935 (0.899–0.971)
T2
0.961 (0.923–0.998)
0.924 (0.835–1.000)
0.916 (0.813–1.000)
0.917 (0.820–1.000)
0.908 (0.837–0.978)
0.895 (0.769–1.000)
0.947 (0.901–0.992)
0.907 (0.779–1.000)
0.890 (0.795–0.985)
T3
0.967 (0.928–1.000)
0.952 (0.904–0.999)
0.942 (0.888–0.997)
0.957 (0.915–1.000)
0.959 (0.919–0.999)
0.938 (0.883–0.992)
0.922 (0.863–0.982)
0.954 (0.909–0.999)
0.967 (0.927–1.000)
T4
0.953 (0.898–0.988)
0.943 (0.868–0.987)
0.950 (0.894–0.988)
0.950 (0.899–0.984)
0.952 (0.893–0.989)
0.944 (0.895–0.980)
0.952 (0.887–0.989)
0.943 (0.875–0.984)
0.949 (0.904–0.980)
T5
0.956 (0.916–0.996)
0.952 (0.907–0.998)
0.952 (0.916–0.988)
0.957 (0.919–0.995)
0.955 (0.906–1.000)
0.951 (0.906–0.995)
0.948 (0.909–0.987)
0.951 (0.912–0.990)
0.909 (0.852–0.966)
T6
0.948 (0.885–1.000)
0.951 (0.892–1.000)
0.941 (0.875–1.000)
0.952 (0.900–1.000)
0.949 (0.888–1.000)
0.937 (0.864–1.000)
0.921 (0.843–0.998)
0.938 (0.871–1.000)
0.897 (0.813–0.981)
Appendix
Extended Data Table 33: Performance across foundation models on diagnostic tasks (T1–T11) in the temporal cohort.
Task
FOCUS
H-optimus-1
H-optimus-0
GigaPath
CRISP
Virchow2
Virchow
UNI
Phikon-v2
T1
0.969 (0.932–1.000)
0.967 (0.937–0.997)
0.964 (0.936–0.993)
0.941 (0.891–0.991)
0.954 (0.910–0.998)
0.956 (0.915–0.998)
0.950 (0.901–0.999)
0.943 (0.889–0.997)
0.895 (0.786–1.000)
T2
0.942 (0.895–0.988)
0.869 (0.760–0.978)
0.858 (0.739–0.977)
0.836 (0.750–0.921)
0.939 (0.900–0.978)
0.842 (0.725–0.958)
0.805 (0.697–0.913)
0.922 (0.874–0.969)
0.814 (0.675–0.954)
T3
0.969 (0.950–0.989)
0.965 (0.944–0.987)
0.961 (0.939–0.983)
0.904 (0.866–0.942)
0.935 (0.903–0.967)
0.971 (0.952–0.990)
0.849 (0.799–0.899)
0.921 (0.886–0.956)
0.833 (0.778–0.888)
T4
0.929 (0.842–0.991)
0.913 (0.790–1.000)
0.904 (0.818–0.972)
0.567 (0.236–0.921)
0.718 (0.495–0.913)
0.603 (0.375–0.840)
0.878 (0.775–0.965)
0.644 (0.457–0.838)
0.638 (0.348–0.895)
T5
0.930 (0.784–1.000)
0.930 (0.844–1.000)
0.870 (0.764–0.976)
0.840 (0.593–1.000)
0.910 (0.726–1.000)
0.920 (0.789–1.000)
0.930 (0.844–1.000)
0.910 (0.817–1.000)
0.900 (0.812–0.988)
T6
0.867 (0.812–0.922)
0.786 (0.715–0.856)
0.819 (0.753–0.885)
0.876 (0.824–0.929)
0.955 (0.928–0.983)
0.754 (0.681–0.827)
0.785 (0.716–0.853)
0.822 (0.759–0.884)
0.670 (0.587–0.753)
Appendix
Extended Data Table 34: Performance across foundation models on diagnostic tasks (T1–T11) in the external cohort.
Task
FOCUS
H-optimus-1
H-optimus-0
GigaPath
CRISP
Virchow2
Virchow
UNI
Phikon-v2
T12
0.841 (0.770–0.913)
0.784 (0.710–0.857)
0.765 (0.684–0.845)
0.763 (0.673–0.852)
0.811 (0.732–0.891)
0.792 (0.709–0.876)
0.739 (0.661–0.816)
0.825 (0.749–0.901)
0.749 (0.669–0.829)
T13
0.908 (0.858–0.959)
0.836 (0.762–0.909)
0.809 (0.734–0.885)
0.884 (0.824–0.944)
0.847 (0.780–0.914)
0.844 (0.775–0.913)
0.829 (0.757–0.900)
0.865 (0.802–0.928)
0.808 (0.733–0.883)
T14
0.847 (0.710–0.985)
0.722 (0.546–0.898)
0.716 (0.517–0.915)
0.808 (0.675–0.941)
0.801 (0.655–0.947)
0.800 (0.683–0.918)
0.762 (0.564–0.960)
0.764 (0.600–0.928)
0.684 (0.520–0.848)
T15
0.923 (0.798–1.000)
0.933 (0.838–1.000)
0.961 (0.911–1.000)
0.946 (0.875–1.000)
0.954 (0.879–1.000)
0.923 (0.817–1.000)
0.928 (0.794–1.000)
0.951 (0.881–1.000)
0.936 (0.880–0.991)
T16
0.886 (0.823–0.948)
0.836 (0.755–0.916)
0.819 (0.689–0.949)
0.850 (0.754–0.946)
0.847 (0.725–0.969)
0.803 (0.675–0.930)
0.785 (0.690–0.879)
0.803 (0.692–0.913)
0.763 (0.638–0.889)
T17
0.872 (0.811–0.933)
0.833 (0.768–0.899)
0.826 (0.755–0.897)
0.830 (0.759–0.901)
0.836 (0.769–0.903)
0.801 (0.725–0.877)
0.827 (0.759–0.895)
0.789 (0.713–0.865)
0.780 (0.704–0.855)
Appendix
Extended Data Table 35: Performance across foundation models on biomarker tasks (T12–T23) in the temporal cohort.
Task
FOCUS
H-optimus-1
H-optimus-0
GigaPath
CRISP
Virchow2
Virchow
UNI
Phikon-v2
T12
1.000 (1.000–1.000)
0.917 (0.773–1.000)
0.979 (0.921–1.000)
0.583 (0.000–1.000)
0.771 (0.467–1.000)
0.896 (0.743–1.000)
0.938 (0.815–1.000)
0.854 (0.556–1.000)
0.979 (0.921–1.000)
T13
1.000 (1.000–1.000)
0.926 (0.798–1.000)
0.926 (0.803–1.000)
0.907 (0.775–1.000)
0.704 (0.219–1.000)
0.722 (0.479–0.965)
1.000 (1.000–1.000)
0.611 (0.319–0.904)
0.593 (0.195–0.991)
T14
0.316 (0.154–0.540)
0.105 (0.029–0.314)
0.158 (0.055–0.376)
0.000 (0.000–0.168)
0.053 (0.009–0.246)
0.000 (0.000–0.168)
0.053 (0.009–0.246)
0.105 (0.029–0.314)
1.000 (0.832–1.000)
T15
1.000 (0.832–1.000)
1.000 (0.832–1.000)
1.000 (0.832–1.000)
1.000 (0.832–1.000)
1.000 (0.832–1.000)
1.000 (0.832–1.000)
1.000 (0.832–1.000)
1.000 (0.832–1.000)
1.000 (0.832–1.000)
T16
1.000 (1.000–1.000)
0.950 (0.842–1.000)
0.850 (0.684–1.000)
0.700 (0.500–0.900)
1.000 (1.000–1.000)
0.950 (0.842–1.000)
0.600 (0.368–0.800)
0.600 (0.400–0.800)
0.500 (0.299–0.737)
T17
1.000 (1.000–1.000)
0.958 (0.867–1.000)
0.979 (0.921–1.000)
0.917 (0.773–1.000)
0.979 (0.921–1.000)
0.896 (0.677–1.000)
0.896 (0.677–1.000)
0.875 (0.617–1.000)
0.917 (0.738–1.000)
Appendix
Extended Data Table 36: Performance across foundation models on biomarker tasks (T12–T23) in the external cohort.
Metric
FOCUS
H-optimus-1
H-optimus-0
GigaPath
CRISP
Virchow2
Virchow
UNI
Phikon-v2
AUC
0.967 (0.936–0.998)
0.943 (0.900–0.986)
0.899 (0.832–0.965)
0.924 (0.869–0.978)
0.935 (0.887–0.982)
0.907 (0.846–0.967)
0.904 (0.832–0.975)
0.956 (0.920–0.992)
0.778 (0.689–0.868)
Accuracy
0.902 (0.829–0.946)
0.892 (0.817–0.939)
0.873 (0.794–0.924)
0.833 (0.749–0.893)
0.892 (0.817–0.939)
0.833 (0.749–0.893)
0.814 (0.727–0.877)
0.843 (0.760–0.901)
0.794 (0.706–0.861)
F1 score
0.940 (0.897–0.972)
0.930 (0.884–0.968)
0.923 (0.876–0.961)
0.904 (0.854–0.946)
0.933 (0.889–0.970)
0.902 (0.848–0.944)
0.869 (0.806–0.921)
0.908 (0.855–0.949)
0.885 (0.834–0.932)
Appendix
Extended Data Table 37: Performance of nine foundation models in the lymphoma differentiation cohort ( n=102 slides; 81 lymphoma and 21 non-lymphoma).
Model
T1 internal
T1 temporal
T1 external
T11 pooled
FOCUS
0.880 (0.800–0.980)
0.862 (0.448–0.966)
0.778 (0.556–1.000)
0.813 (0.689–0.957)
H-Optimus-1
0.833 (0.741–0.944)
0.759 (0.552–0.897)
0.778 (0.444–1.000)
0.680 (0.571–0.861)
H-Optimus-0
0.833 (0.741–0.944)
0.759 (0.586–0.931)
0.667 (0.333–1.000)
0.688 (0.599–0.860)
GigaPath
0.815 (0.704–0.907)
0.724 (0.517–0.897)
0.667 (0.333–1.000)
0.803 (0.538–0.942)
CRISP
0.796 (0.667–0.907)
0.621 (0.448–0.862)
0.778 (0.444–1.000)
0.724 (0.536–0.930)
Virchow2
0.833 (0.722–0.926)
0.724 (0.517–0.931)
0.667 (0.333–1.000)
0.786 (0.638–0.917)
Appendix
Extended Data Table 38: Specificity at 95% sensitivity for T1 across three cohorts and T11 in pooled cohorts.
Model
Level 1
Level 2
Level 3
Level 4
FOCUS
0.933 (0.911–0.951)
0.931 (0.912–0.947)
0.905 (0.884–0.921)
0.865 (0.845–0.882)
H-optimus-1
0.922 (0.901–0.942)
0.923 (0.902–0.940)
0.891 (0.871–0.909)
0.848 (0.828–0.866)
H-optimus-0
0.922 (0.898–0.942)
0.921 (0.900–0.940)
0.887 (0.867–0.907)
0.844 (0.823–0.864)
GigaPath
0.911 (0.887–0.933)
0.913 (0.890–0.930)
0.881 (0.859–0.899)
0.842 (0.820–0.861)
CRISP
0.899 (0.878–0.924)
0.907 (0.887–0.927)
0.871 (0.850–0.891)
0.830 (0.810–0.850)
Virchow2
0.937 (0.916–0.956)
0.934 (0.915–0.950)
0.901 (0.881–0.918)
0.858 (0.839–0.876)
Appendix
Extended Data Table 39: Level-wise H-F1 for the Test dataset.
Model
Level 1
Level 2
Level 3
Level 4
FOCUS
0.933 (0.911–0.951)
0.931 (0.911–0.947)
0.890 (0.869–0.908)
0.825 (0.801–0.847)
H-optimus-1
0.922 (0.901–0.942)
0.923 (0.903–0.940)
0.873 (0.852–0.893)
0.804 (0.781–0.826)
H-optimus-0
0.922 (0.898–0.942)
0.921 (0.900–0.939)
0.869 (0.846–0.891)
0.800 (0.776–0.824)
GigaPath
0.911 (0.887–0.933)
0.913 (0.891–0.931)
0.864 (0.840–0.884)
0.803 (0.778–0.823)
CRISP
0.899 (0.878–0.924)
0.909 (0.890–0.929)
0.853 (0.830–0.875)
0.788 (0.765–0.810)
Virchow2
0.937 (0.916–0.956)
0.932 (0.913–0.949)
0.882 (0.860–0.901)
0.814 (0.791–0.836)
Appendix
Extended Data Table 40: Level-wise depth-weighted H-F1 (DH-F1) for the Test dataset.
Model
Level 1
Level 2
Level 3
Level 4
FOCUS
0.933 (0.911–0.951)
0.905 (0.878–0.927)
0.830 (0.800–0.856)
0.702 (0.667–0.736)
H-optimus-1
0.922 (0.901–0.942)
0.895 (0.867–0.917)
0.804 (0.772–0.832)
0.668 (0.632–0.705)
H-optimus-0
0.922 (0.898–0.942)
0.891 (0.864–0.916)
0.798 (0.765–0.830)
0.659 (0.619–0.697)
GigaPath
0.911 (0.887–0.933)
0.881 (0.852–0.905)
0.789 (0.757–0.818)
0.665 (0.627–0.699)
CRISP
0.899 (0.878–0.924)
0.870 (0.844–0.898)
0.757 (0.722–0.789)
0.610 (0.570–0.647)
Virchow2
0.937 (0.916–0.956)
0.907 (0.881–0.930)
0.817 (0.784–0.846)
0.679 (0.642–0.717)
Appendix
Extended Data Table 41: Level-wise Accuracy for the Test dataset.
Model
Level 1
Level 2
Level 3
Level 4
FOCUS
0.926 (0.893–0.955)
0.904 (0.874–0.930)
0.863 (0.831–0.893)
0.816 (0.785–0.845)
H-optimus-1
0.893 (0.858–0.926)
0.864 (0.832–0.894)
0.820 (0.785–0.853)
0.764 (0.731–0.797)
H-optimus-0
0.909 (0.877–0.942)
0.888 (0.858–0.918)
0.850 (0.817–0.882)
0.787 (0.754–0.818)
GigaPath
0.867 (0.829–0.903)
0.853 (0.818–0.885)
0.814 (0.779–0.847)
0.757 (0.721–0.790)
CRISP
0.884 (0.848–0.916)
0.880 (0.849–0.909)
0.842 (0.811–0.871)
0.787 (0.759–0.816)
Virchow2
0.906 (0.874–0.935)
0.862 (0.832–0.893)
0.795 (0.761–0.826)
0.754 (0.719–0.786)
Appendix
Extended Data Table 42: Level-wise H-F1 for the Temporal Validation Cohort.
Model
Level 1
Level 2
Level 3
Level 4
FOCUS
0.926 (0.893–0.955)
0.896 (0.864–0.924)
0.836 (0.799–0.870)
0.761 (0.724–0.796)
H-optimus-1
0.893 (0.858–0.926)
0.853 (0.820–0.886)
0.790 (0.751–0.828)
0.699 (0.662–0.738)
H-optimus-0
0.909 (0.877–0.942)
0.881 (0.849–0.912)
0.825 (0.786–0.861)
0.721 (0.682–0.758)
GigaPath
0.867 (0.829–0.903)
0.848 (0.812–0.882)
0.789 (0.749–0.827)
0.694 (0.653–0.731)
CRISP
0.884 (0.848–0.916)
0.879 (0.847–0.909)
0.822 (0.787–0.855)
0.731 (0.696–0.766)
Virchow2
0.906 (0.874–0.935)
0.846 (0.813–0.879)
0.749 (0.707–0.787)
0.688 (0.647–0.728)
Appendix
Extended Data Table 43: Level-wise depth-weighted DH-F1 for the Temporal Validation Cohort.
Model
Level 1
Level 2
Level 3
Level 4
FOCUS
0.926 (0.893–0.955)
0.854 (0.809–0.893)
0.738 (0.689–0.786)
0.576 (0.521–0.631)
H-optimus-1
0.893 (0.858–0.926)
0.799 (0.754–0.845)
0.683 (0.631–0.731)
0.498 (0.443–0.553)
H-optimus-0
0.909 (0.877–0.942)
0.835 (0.790–0.877)
0.725 (0.670–0.777)
0.518 (0.460–0.576)
GigaPath
0.867 (0.829–0.903)
0.793 (0.748–0.838)
0.676 (0.625–0.728)
0.492 (0.434–0.547)
CRISP
0.884 (0.848–0.916)
0.825 (0.780–0.867)
0.696 (0.647–0.748)
0.485 (0.430–0.541)
Virchow2
0.906 (0.874–0.935)
0.790 (0.744–0.832)
0.625 (0.570–0.676)
0.518 (0.463–0.573)
Appendix
Extended Data Table 44: Level-wise Accuracy for the Temporal Validation Cohort.
Model
Level 1
Level 2
Level 3
Level 4
FOCUS
0.911 (0.874–0.943)
0.895 (0.862–0.926)
0.825 (0.791–0.857)
0.772 (0.737–0.805)
H-optimus-1
0.915 (0.878–0.947)
0.826 (0.791–0.859)
0.706 (0.669–0.741)
0.680 (0.640–0.717)
H-optimus-0
0.911 (0.874–0.943)
0.877 (0.847–0.913)
0.781 (0.747–0.813)
0.729 (0.696–0.763)
GigaPath
0.911 (0.874–0.943)
0.855 (0.821–0.887)
0.733 (0.696–0.766)
0.687 (0.651–0.719)
CRISP
0.911 (0.874–0.943)
0.865 (0.831–0.897)
0.787 (0.752–0.821)
0.743 (0.708–0.777)
Virchow2
0.911 (0.874–0.943)
0.869 (0.837–0.901)
0.717 (0.684–0.749)
0.679 (0.646–0.711)
Appendix
Extended Data Table 45: Level-wise H-F1 for the External Validation Cohort.
Model
Level 1
Level 2
Level 3
Level 4
FOCUS
0.911 (0.874–0.943)
0.889 (0.856–0.922)
0.784 (0.745–0.819)
0.690 (0.650–0.729)
H-optimus-1
0.915 (0.878–0.947)
0.796 (0.756–0.834)
0.624 (0.579–0.666)
0.578 (0.531–0.628)
H-optimus-0
0.911 (0.874–0.943)
0.866 (0.834–0.903)
0.722 (0.681–0.760)
0.632 (0.593–0.670)
GigaPath
0.911 (0.874–0.943)
0.835 (0.798–0.872)
0.656 (0.613–0.695)
0.580 (0.538–0.618)
CRISP
0.911 (0.874–0.943)
0.849 (0.812–0.884)
0.736 (0.694–0.777)
0.659 (0.617–0.699)
Virchow2
0.911 (0.874–0.943)
0.855 (0.820–0.890)
0.626 (0.587–0.664)
0.563 (0.524–0.599)
Appendix
Extended Data Table 46: Level-wise DH-F1 for the External Validation Cohort.
Model
Level 1
Level 2
Level 3
Level 4
FOCUS
0.911 (0.874–0.943)
0.837 (0.789–0.886)
0.553 (0.484–0.610)
0.411 (0.346–0.468)
H-optimus-1
0.915 (0.878–0.947)
0.703 (0.646–0.756)
0.358 (0.297–0.419)
0.346 (0.289–0.402)
H-optimus-0
0.911 (0.874–0.943)
0.805 (0.760–0.858)
0.463 (0.402–0.524)
0.325 (0.268–0.382)
GigaPath
0.911 (0.874–0.943)
0.760 (0.707–0.813)
0.374 (0.313–0.431)
0.264 (0.207–0.317)
CRISP
0.911 (0.874–0.943)
0.781 (0.728–0.829)
0.504 (0.443–0.565)
0.382 (0.321–0.439)
Virchow2
0.911 (0.874–0.943)
0.789 (0.740–0.837)
0.301 (0.244–0.362)
0.203 (0.155–0.260)
Appendix
Extended Data Table 47: Level-wise Accuracy for the External Validation Cohort.
Model
Level 1
Level 2
Level 3
Level 4
FOCUS
0.909 (0.844–0.961)
0.902 (0.844–0.957)
0.857 (0.797–0.915)
0.813 (0.753–0.873)
H-optimus-1
0.909 (0.844–0.974)
0.905 (0.846–0.960)
0.842 (0.784–0.903)
0.784 (0.726–0.840)
H-optimus-0
0.909 (0.844–0.961)
0.899 (0.841–0.951)
0.825 (0.768–0.884)
0.761 (0.705–0.818)
GigaPath
0.922 (0.857–0.974)
0.916 (0.860–0.967)
0.835 (0.780–0.891)
0.785 (0.730–0.840)
CRISP
0.961 (0.909–1.000)
0.944 (0.902–0.981)
0.856 (0.807–0.904)
0.808 (0.761–0.858)
Virchow2
0.935 (0.870–0.987)
0.910 (0.856–0.957)
0.831 (0.772–0.887)
0.777 (0.715–0.831)
Appendix
Extended Data Table 48: Level-wise H-F1 for the Uncertainty Group.
Model
Level 1
Level 2
Level 3
Level 4
FOCUS
0.909 (0.844–0.961)
0.900 (0.840–0.956)
0.831 (0.765–0.899)
0.761 (0.689–0.831)
H-optimus-1
0.909 (0.844–0.974)
0.904 (0.844–0.960)
0.808 (0.744–0.878)
0.715 (0.647–0.783)
H-optimus-0
0.909 (0.844–0.961)
0.896 (0.837–0.947)
0.785 (0.717–0.857)
0.682 (0.615–0.748)
GigaPath
0.922 (0.857–0.974)
0.914 (0.858–0.965)
0.792 (0.728–0.858)
0.715 (0.646–0.780)
CRISP
0.961 (0.909–1.000)
0.938 (0.891–0.982)
0.808 (0.748–0.869)
0.736 (0.672–0.801)
Virchow2
0.935 (0.870–0.987)
0.901 (0.840–0.955)
0.785 (0.719–0.854)
0.701 (0.632–0.768)
Appendix
Extended Data Table 49: Level-wise DH-F1 for the Uncertainty Group.
Model
Level 1
Level 2
Level 3
Level 4
FOCUS
0.909 (0.844–0.961)
0.857 (0.779–0.935)
0.727 (0.636–0.831)
0.623 (0.520–0.727)
H-optimus-1
0.909 (0.844–0.974)
0.857 (0.766–0.935)
0.675 (0.571–0.779)
0.520 (0.416–0.623)
H-optimus-0
0.909 (0.844–0.961)
0.844 (0.753–0.922)
0.636 (0.533–0.753)
0.455 (0.351–0.558)
GigaPath
0.922 (0.857–0.974)
0.870 (0.792–0.948)
0.636 (0.533–0.740)
0.507 (0.403–0.610)
CRISP
0.961 (0.909–1.000)
0.909 (0.844–0.974)
0.662 (0.558–0.754)
0.546 (0.442–0.649)
Virchow2
0.935 (0.870–0.987)
0.857 (0.779–0.935)
0.649 (0.546–0.753)
0.520 (0.416–0.623)
Appendix
Extended Data Table 50: Level-wise Accuracy for the Uncertainty Group.
Model
Level 1
Level 2
Level 3
Level 4
FOCUS
0.924 (0.864–0.985)
0.879 (0.819–0.935)
0.723 (0.661–0.781)
0.637 (0.583–0.691)
H-optimus-1
0.697 (0.576–0.803)
0.697 (0.587–0.792)
0.566 (0.482–0.646)
0.498 (0.421–0.569)
H-optimus-0
0.712 (0.591–0.818)
0.703 (0.603–0.797)
0.568 (0.486–0.650)
0.506 (0.433–0.582)
GigaPath
0.697 (0.576–0.818)
0.700 (0.592–0.795)
0.571 (0.485–0.652)
0.510 (0.433–0.584)
CRISP
0.682 (0.561–0.788)
0.675 (0.564–0.771)
0.552 (0.467–0.634)
0.490 (0.411–0.566)
Virchow2
0.773 (0.667–0.879)
0.755 (0.661–0.844)
0.639 (0.551–0.721)
0.562 (0.486–0.634)
Appendix
Extended Data Table 51: Level-wise H-F1 for the Discordance Group.
Model
Level 1
Level 2
Level 3
Level 4
FOCUS
0.924 (0.864–0.985)
0.864 (0.797–0.927)
0.634 (0.561–0.705)
0.504 (0.441–0.568)
H-optimus-1
0.697 (0.576–0.803)
0.697 (0.587–0.791)
0.493 (0.414–0.578)
0.386 (0.321–0.456)
H-optimus-0
0.712 (0.591–0.818)
0.699 (0.598–0.795)
0.493 (0.411–0.575)
0.399 (0.329–0.476)
GigaPath
0.697 (0.576–0.818)
0.701 (0.595–0.794)
0.501 (0.420–0.583)
0.408 (0.335–0.484)
CRISP
0.682 (0.561–0.788)
0.672 (0.560–0.769)
0.482 (0.399–0.565)
0.386 (0.315–0.464)
Virchow2
0.773 (0.667–0.879)
0.749 (0.653–0.840)
0.575 (0.486–0.664)
0.450 (0.382–0.523)
Appendix
Extended Data Table 52: Level-wise DH-F1 for the Discordance Group.
Model
Level 1
Level 2
Level 3
Level 4
FOCUS
0.924 (0.864–0.985)
0.803 (0.712–0.894)
0.394 (0.288–0.500)
0.227 (0.136–0.333)
H-optimus-1
0.697 (0.576–0.803)
0.591 (0.470–0.712)
0.242 (0.152–0.349)
0.121 (0.061–0.212)
H-optimus-0
0.712 (0.591–0.818)
0.591 (0.470–0.712)
0.242 (0.152–0.349)
0.136 (0.061–0.227)
GigaPath
0.697 (0.576–0.818)
0.591 (0.469–0.712)
0.242 (0.152–0.349)
0.136 (0.061–0.227)
CRISP
0.682 (0.561–0.788)
0.561 (0.439–0.682)
0.242 (0.152–0.349)
0.152 (0.076–0.243)
Virchow2
0.773 (0.667–0.879)
0.652 (0.530–0.773)
0.349 (0.242–0.470)
0.182 (0.106–0.288)
Appendix
Extended Data Table 53: Level-wise Accuracy for the Discordance Group.
Task
Diagnostic task
Without clinical
With clinical
Δ
T9
Diffuse glioma molecular-subtype classification
0.928 (0.896–0.952)
0.956 (0.940–0.971)
+0.028
T5
Glioma vs GNT
0.899 (0.862–0.936)
0.914 (0.883–0.947)
+0.016
T8
Diffuse vs circumscribed glioma
0.956 (0.931–0.977)
0.971 (0.953–0.985)
+0.015
T7
Non-neuroepithelial tumour classification
0.978 (0.949–0.997)
0.985 (0.959–0.999)
+0.008
T1
Non-neoplastic brain tissue vs tumour
0.964 (0.941–0.983)
0.972 (0.954–0.987)
+0.008
T10
PA vs PXA
0.988 (0.958–1.000)
0.990 (0.965–1.000)
+0.002
Appendix
Extended Data Table 54: Macro AUC before and after incorporating clinical information for the Tiantan test-set diagnostic tasks.
Metric
Without clinical
With clinical
Δ
Hierarchical F1 score
0.865 (0.845–0.882)
0.888 (0.871–0.905)
+0.024
Depth-weighted hierarchical F1 score
0.825 (0.801–0.847)
0.854 (0.833–0.876)
+0.029
Leaf accuracy
0.702 (0.667–0.736)
0.755 (0.720–0.791)
+0.054
Balanced accuracy
0.578 (0.528–0.627)
0.628 (0.580–0.678)
+0.050
Leaf macro F1 score
0.582 (0.530–0.623)
0.633 (0.578–0.676)
+0.051
Leaf weighted F1 score
0.691 (0.655–0.728)
0.746 (0.710–0.781)
+0.055
Appendix
Extended Data Table 55: Hierarchical and leaf-level classification metrics before and after incorporating clinical information on the Tiantan test set.
Reader
H-F1
H-F1 + AI
Δ H-F1
P
DH-F1
DH-F1 + AI
Δ DH-F1
P
Junior 1
0.735
0.796
+0.061
<0.001
0.645
0.716
+0.070
0.003
Junior 2
0.670
0.817
+0.147
<0.001
0.556
0.724
+0.168
<0.001
Junior 3
0.682
0.746
+0.064
<0.001
0.580
0.649
+0.069
<0.001
Senior 1
0.785
0.803
+0.017
0.031
0.690
0.712
+0.022
0.031
Senior 2
0.776
0.862
+0.086
<0.001
0.680
0.799
+0.119
<0.001
Senior 3
0.876
0.876
+0.000
1.000
0.810
0.810
+0.000
1.000
Appendix
Extended Data Table 56: Paired hierarchical performance in the Randomly Selected Group.
Reader
H-F1
H-F1 + AI
Δ H-F1
P
DH-F1
DH-F1 + AI
Δ DH-F1
P
Junior 1
0.711
0.784
+0.073
0.001
0.593
0.680
+0.087
0.002
Junior 2
0.654
0.707
+0.053
0.039
0.529
0.593
+0.064
0.024
Junior 3
0.767
0.783
+0.016
0.299
0.677
0.698
+0.021
0.172
Senior 1
0.719
0.767
+0.048
0.004
0.629
0.686
+0.057
0.003
Senior 2
0.702
0.807
+0.105
<0.001
0.587
0.723
+0.136
<0.001
Senior 3
0.817
0.834
+0.017
0.064
0.754
0.774
+0.020
0.064
Appendix
Extended Data Table 57: Paired hierarchical performance in the Uncertainty Group.
Reader
H-F1
H-F1 + AI
Δ H-F1
P
DH-F1
DH-F1 + AI
Δ DH-F1
P
Junior 1
0.670
0.682
+0.012
0.408
0.553
0.566
+0.013
0.433
Junior 2
0.600
0.591
-0.009
0.932
0.463
0.451
-0.012
0.931
Junior 3
0.419
0.476
+0.057
0.010
0.326
0.387
+0.061
0.010
Senior 1
0.496
0.508
+0.012
0.125
0.393
0.404
+0.011
0.125
Senior 2
0.466
0.568
+0.102
0.013
0.360
0.468
+0.108
0.013
Senior 3
0.529
0.550
+0.021
0.126
0.462
0.477
+0.015
0.126
Appendix
Extended Data Table 58: Paired hierarchical performance in the Discordance Group.
Reader
n
Unaided
AI-assisted
Mean Δ
SD
95% CI
Junior 1
100
2.14
2.36
+0.22
0.61
[0.10, 0.34]
Junior 2
100
2.08
2.54
+0.46
0.86
[0.29, 0.63]
Junior 3
100
2.20
2.35
+0.15
0.44
[0.06, 0.24]
Senior 1
100
2.36
2.42
+0.06
0.34
[-0.01, 0.13]
Senior 2
100
2.27
2.71
+0.44
0.69
[0.31, 0.57]
Senior 3
100
2.61
2.61
+0.00
0.00
[0.00, 0.00]
Appendix
Extended Data Table 59: Paired maximum correct depth in the Randomly Selected Group.
Reader
n
Unaided
AI-assisted
Mean Δ
SD
95% CI
Junior 1
77
2.16
2.38
+0.22
0.60
[0.09, 0.35]
Junior 2
77
2.16
2.38
+0.22
0.91
[0.02, 0.42]
Junior 3
77
2.38
2.44
+0.06
0.34
[-0.01, 0.14]
Senior 1
77
2.19
2.39
+0.19
0.56
[0.07, 0.32]
Senior 2
77
2.06
2.56
+0.49
0.75
[0.33, 0.66]
Senior 3
77
2.52
2.57
+0.05
0.22
[0.00, 0.10]
Appendix
Extended Data Table 60: Paired maximum correct depth in the Uncertainty Group.
Reader
n
Unaided
AI-assisted
Mean Δ
SD
95% CI
Junior 1
66
2.02
2.06
+0.05
0.27
[-0.02, 0.11]
Junior 2
66
1.86
2.00
+0.14
0.96
[-0.09, 0.37]
Junior 3
66
1.14
1.29
+0.15
0.44
[0.05, 0.26]
Senior 1
66
1.44
1.47
+0.03
0.25
[-0.03, 0.09]
Senior 2
66
1.24
1.62
+0.38
1.03
[0.13, 0.63]
Senior 3
66
1.36
1.41
+0.05
0.21
[-0.01, 0.10]
Appendix
Extended Data Table 61: Paired maximum correct depth in the Discordance Group.
Reader
Level 1
Level 2
Level 3
Level 4
Junior 1
0.899/0.899 (+0.000)
0.747/0.848 (+0.101)
0.545/0.667 (+0.121)
0.232/0.313 (+0.081)
Junior 2
0.876/0.980 (+0.104)
0.784/0.930 (+0.146)
0.485/0.760 (+0.275)
0.134/0.490 (+0.356)
Junior 3
0.879/0.920 (+0.041)
0.758/0.830 (+0.072)
0.566/0.680 (+0.114)
0.222/0.340 (+0.118)
Senior 1
0.940/0.940 (+0.000)
0.860/0.890 (+0.030)
0.690/0.720 (+0.030)
0.370/0.400 (+0.030)
Senior 2
0.949/0.980 (+0.031)
0.838/0.940 (+0.102)
0.667/0.790 (+0.123)
0.273/0.610 (+0.337)
Senior 3
0.980/0.980 (+0.000)
0.960/0.960 (+0.000)
0.800/0.800 (+0.000)
0.550/0.550 (+0.000)
Appendix
Extended Data Table 62: Level-wise leaf accuracy in the Randomly Selected Group.
Reader
Level 1
Level 2
Level 3
Level 4
Junior 1
0.948/0.961 (+0.013)
0.792/0.883 (+0.091)
0.416/0.558 (+0.143)
0.143/0.273 (+0.130)
Junior 2
0.932/0.922 (-0.009)
0.767/0.831 (+0.064)
0.521/0.597 (+0.077)
0.205/0.312 (+0.106)
Junior 3
0.935/0.935 (+0.000)
0.831/0.844 (+0.013)
0.597/0.649 (+0.052)
0.312/0.351 (+0.039)
Senior 1
0.921/0.921 (+0.000)
0.684/0.789 (+0.105)
0.553/0.645 (+0.092)
0.316/0.382 (+0.066)
Senior 2
0.933/0.947 (+0.014)
0.747/0.895 (+0.148)
0.467/0.724 (+0.257)
0.173/0.421 (+0.248)
Senior 3
0.921/0.921 (+0.000)
0.842/0.868 (+0.026)
0.711/0.737 (+0.026)
0.434/0.461 (+0.026)
Appendix
Extended Data Table 63: Level-wise leaf accuracy in the Uncertainty Group.
Reader
Level 1
Level 2
Level 3
Level 4
Junior 1
0.909/0.909 (+0.000)
0.773/0.788 (+0.015)
0.333/0.364 (+0.030)
0.121/0.167 (+0.045)
Junior 2
0.898/0.894 (-0.004)
0.780/0.758 (-0.022)
0.356/0.379 (+0.023)
0.102/0.136 (+0.035)
Junior 3
0.591/0.621 (+0.030)
0.439/0.530 (+0.091)
0.182/0.273 (+0.091)
0.167/0.227 (+0.061)
Senior 1
0.677/0.692 (+0.015)
0.554/0.569 (+0.015)
0.215/0.215 (+0.000)
0.092/0.092 (+0.000)
Senior 2
0.662/0.742 (+0.081)
0.477/0.591 (+0.114)
0.123/0.288 (+0.165)
0.046/0.121 (+0.075)
Senior 3
0.636/0.667 (+0.030)
0.470/0.515 (+0.045)
0.333/0.364 (+0.030)
0.288/0.318 (+0.030)
Appendix
Extended Data Table 64: Level-wise leaf accuracy in the Discordance Group.
Reader
Accuracy
Accuracy + AI
Δ Accuracy
P
Lymphoma recall
Recall + AI
Δ Recall
P
Junior 1
0.735
0.843
+0.108
0.090
0.753
0.988
+0.235
<0.001
Junior 2
0.660
0.718
+0.058
0.164
0.780
0.866
+0.085
0.062
Junior 3
0.843
0.941
+0.098
0.062
0.914
0.988
+0.074
0.062
Junior mean
0.746
0.834
+0.088
<0.001
0.816
0.947
+0.131
<0.001
Senior 1
0.863
0.892
+0.029
0.333
0.864
0.889
+0.025
0.500
Senior 2
0.725
0.784
+0.059
0.338
0.753
0.877
+0.123
0.062
Appendix
Extended Data Table 65: Paired reader performance in the lymphoma differentiation cohort.
Hardware configuration
Preprocessing (s)
Prediction (s)
Total (s)
Intel(R) Xeon(R) Platinum 8468V CPU + NVIDIA A100 80GB GPU
96.47
0.97
97.44
AMD EPYC 7K62 CPU + NVIDIA A6000 48GB GPU
110.71
0.89
111.60
Intel(R) Xeon(R) Platinum 8157 CPU + NVIDIA RTX 3090 GPU
138.47
1.00
139.47
Appendix
Extended Data Table 66: FOCUS processing time across tested hardware configurations.
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
GigaPath
1.000 (1.000–1.000)
0.993 (0.979–1.000)
1.000 (1.000–1.000)
0.909 (0.700–1.000)
0.996 (0.988–1.000)
H-optimus-0
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
H-optimus-1
1.000 (1.000–1.000)
0.993 (0.979–1.000)
1.000 (1.000–1.000)
0.909 (0.700–1.000)
0.996 (0.988–1.000)
Phikon-v2
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
UNI
1.000 (1.000–1.000)
0.993 (0.979–1.000)
1.000 (1.000–1.000)
0.909 (0.700–1.000)
0.996 (0.988–1.000)
Appendix
Extended Data Table 67: Test-set metrics for BLCA (Benign vs. malignant).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
1.000 (1.000–1.000)
0.998 (0.995–1.000)
0.998 (0.993–1.000)
1.000 (1.000–1.000)
0.999 (0.997–1.000)
GigaPath
1.000 (1.000–1.000)
0.995 (0.988–1.000)
0.994 (0.985–1.000)
1.000 (1.000–1.000)
0.997 (0.993–1.000)
H-optimus-0
1.000 (1.000–1.000)
0.997 (0.992–1.000)
0.996 (0.989–1.000)
1.000 (1.000–1.000)
0.998 (0.994–1.000)
H-optimus-1
1.000 (1.000–1.000)
0.997 (0.992–1.000)
0.996 (0.989–1.000)
1.000 (1.000–1.000)
0.998 (0.994–1.000)
Phikon-v2
1.000 (1.000–1.000)
0.997 (0.992–1.000)
0.996 (0.989–1.000)
1.000 (1.000–1.000)
0.998 (0.994–1.000)
UNI
1.000 (1.000–1.000)
0.998 (0.995–1.000)
0.998 (0.993–1.000)
1.000 (1.000–1.000)
0.999 (0.997–1.000)
Appendix
Extended Data Table 68: Test-set metrics for BRCA (Benign vs. malignant).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
GigaPath
0.773 (0.708–0.829)
0.620 (0.553–0.683)
0.428 (0.392–0.462)
0.752 (0.719–0.783)
0.568 (0.487–0.642)
H-optimus-0
0.721 (0.656–0.783)
0.591 (0.524–0.659)
0.408 (0.370–0.448)
0.735 (0.699–0.770)
0.547 (0.468–0.622)
H-optimus-1
0.760 (0.699–0.820)
0.611 (0.543–0.678)
0.422 (0.385–0.458)
0.747 (0.713–0.781)
0.565 (0.488–0.639)
Phikon-v2
0.709 (0.646–0.769)
0.601 (0.534–0.668)
0.415 (0.379–0.450)
0.740 (0.708–0.773)
0.549 (0.471–0.623)
UNI
0.767 (0.709–0.824)
0.639 (0.572–0.702)
0.443 (0.404–0.480)
0.765 (0.730–0.800)
0.602 (0.523–0.674)
Virchow
0.805 (0.751–0.857)
0.707 (0.639–0.764)
0.493 (0.451–0.535)
0.812 (0.773–0.848)
0.688 (0.619–0.750)
Appendix
Extended Data Table 69: External-set metrics for CESC (Tumor grading).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.969 (0.945–0.993)
0.848 (0.794–0.897)
0.977 (0.949–1.000)
0.333 (0.174–0.486)
0.912 (0.875–0.943)
GigaPath
0.947 (0.914–0.979)
0.836 (0.782–0.891)
0.970 (0.937–0.993)
0.303 (0.152–0.464)
0.905 (0.867–0.938)
H-optimus-0
0.959 (0.930–0.987)
0.842 (0.782–0.897)
0.985 (0.962–1.000)
0.273 (0.132–0.429)
0.909 (0.871–0.942)
H-optimus-1
0.962 (0.935–0.989)
0.848 (0.794–0.897)
0.977 (0.949–1.000)
0.333 (0.174–0.486)
0.912 (0.875–0.943)
Phikon-v2
0.965 (0.938–0.991)
0.824 (0.764–0.879)
0.992 (0.977–1.000)
0.152 (0.036–0.289)
0.900 (0.862–0.934)
UNI
0.965 (0.939–0.992)
0.842 (0.788–0.891)
0.970 (0.937–0.993)
0.333 (0.174–0.486)
0.908 (0.871–0.940)
Appendix
Extended Data Table 70: External-set metrics for COADREAD (Benign vs. malignant).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.978 (0.952–1.000)
0.962 (0.939–0.985)
0.994 (0.981–1.000)
0.902 (0.840–0.957)
0.971 (0.953–0.988)
GigaPath
0.977 (0.948–1.000)
0.970 (0.947–0.989)
0.982 (0.960–1.000)
0.946 (0.894–0.989)
0.977 (0.959–0.991)
H-optimus-0
0.980 (0.955–1.000)
0.962 (0.939–0.985)
0.977 (0.952–0.994)
0.935 (0.884–0.980)
0.971 (0.952–0.988)
H-optimus-1
0.983 (0.960–1.000)
0.985 (0.970–0.996)
0.994 (0.981–1.000)
0.967 (0.928–1.000)
0.988 (0.975–0.997)
Phikon-v2
0.976 (0.954–0.998)
0.654 (0.593–0.711)
1.000 (1.000–1.000)
0.011 (0.000–0.035)
0.790 (0.745–0.829)
UNI
0.954 (0.923–0.984)
0.821 (0.776–0.867)
0.965 (0.935–0.989)
0.554 (0.452–0.655)
0.875 (0.840–0.909)
Appendix
Extended Data Table 71: External-set metrics for HNSC (Benign vs. malignant).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.634 (0.553–0.709)
0.462 (0.370–0.555)
0.290 (0.229–0.358)
0.764 (0.729–0.799)
0.430 (0.333–0.525)
GigaPath
0.606 (0.527–0.686)
0.445 (0.353–0.538)
0.264 (0.214–0.321)
0.752 (0.719–0.786)
0.397 (0.301–0.491)
H-optimus-0
0.612 (0.530–0.688)
0.504 (0.412–0.588)
0.284 (0.242–0.328)
0.771 (0.743–0.800)
0.424 (0.328–0.522)
H-optimus-1
0.629 (0.552–0.706)
0.471 (0.378–0.555)
0.306 (0.246–0.374)
0.773 (0.738–0.810)
0.444 (0.347–0.539)
Phikon-v2
0.568 (0.484–0.646)
0.454 (0.361–0.538)
0.259 (0.214–0.307)
0.755 (0.723–0.787)
0.391 (0.297–0.484)
UNI
0.621 (0.544–0.698)
0.445 (0.353–0.538)
0.262 (0.212–0.319)
0.751 (0.720–0.784)
0.393 (0.295–0.494)
Appendix
Extended Data Table 72: Test-set metrics for LIHC (Tumor grading).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
GigaPath
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
H-optimus-0
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
H-optimus-1
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
Phikon-v2
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
UNI
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
Appendix
Extended Data Table 73: Test-set metrics for LIHC (Benign vs. malignant).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.999 (0.998–1.000)
0.995 (0.988–1.000)
0.994 (0.986–1.000)
1.000 (1.000–1.000)
0.997 (0.993–1.000)
GigaPath
0.998 (0.995–1.000)
0.990 (0.981–0.998)
0.992 (0.981–1.000)
0.980 (0.933–1.000)
0.994 (0.989–0.999)
H-optimus-0
0.999 (0.996–1.000)
0.993 (0.983–1.000)
0.994 (0.986–1.000)
0.980 (0.933–1.000)
0.996 (0.990–1.000)
H-optimus-1
0.999 (0.997–1.000)
0.995 (0.988–1.000)
0.994 (0.986–1.000)
1.000 (1.000–1.000)
0.997 (0.993–1.000)
Phikon-v2
0.998 (0.995–1.000)
0.993 (0.983–1.000)
0.994 (0.986–1.000)
0.980 (0.933–1.000)
0.996 (0.990–1.000)
UNI
0.998 (0.994–1.000)
0.990 (0.981–0.998)
0.992 (0.981–1.000)
0.980 (0.933–1.000)
0.994 (0.989–0.999)
Appendix
Extended Data Table 74: Test-set metrics for OV (Benign vs. malignant).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.968 (0.949–0.987)
0.908 (0.880–0.936)
0.959 (0.934–0.981)
0.803 (0.734–0.871)
0.934 (0.912–0.955)
GigaPath
0.958 (0.934–0.982)
0.891 (0.860–0.919)
0.985 (0.969–0.996)
0.693 (0.610–0.772)
0.924 (0.901–0.946)
H-optimus-0
0.956 (0.932–0.979)
0.906 (0.875–0.934)
0.940 (0.909–0.968)
0.835 (0.767–0.898)
0.931 (0.908–0.953)
H-optimus-1
0.962 (0.941–0.983)
0.916 (0.885–0.941)
0.917 (0.883–0.949)
0.913 (0.859–0.959)
0.937 (0.913–0.957)
Phikon-v2
0.885 (0.849–0.920)
0.684 (0.639–0.728)
1.000 (1.000–1.000)
0.024 (0.000–0.052)
0.811 (0.777–0.842)
UNI
0.923 (0.894–0.952)
0.781 (0.740–0.822)
0.981 (0.964–0.996)
0.362 (0.279–0.450)
0.859 (0.828–0.886)
Appendix
Extended Data Table 75: External-set metrics for PAAD (Benign vs. malignant).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.774 (0.731–0.818)
0.505 (0.434–0.577)
0.393 (0.331–0.455)
0.867 (0.848–0.885)
0.476 (0.399–0.551)
GigaPath
0.743 (0.693–0.792)
0.467 (0.396–0.538)
0.338 (0.291–0.385)
0.856 (0.838–0.873)
0.417 (0.338–0.493)
H-optimus-0
0.730 (0.681–0.778)
0.478 (0.407–0.549)
0.333 (0.296–0.371)
0.857 (0.840–0.874)
0.400 (0.320–0.477)
H-optimus-1
0.744 (0.696–0.792)
0.489 (0.418–0.560)
0.348 (0.306–0.391)
0.859 (0.842–0.876)
0.415 (0.335–0.494)
Phikon-v2
0.689 (0.634–0.742)
0.451 (0.379–0.522)
0.309 (0.273–0.346)
0.845 (0.830–0.862)
0.356 (0.283–0.434)
UNI
0.752 (0.704–0.800)
0.505 (0.434–0.577)
0.388 (0.331–0.446)
0.865 (0.847–0.884)
0.470 (0.395–0.546)
Appendix
Extended Data Table 76: Test-set metrics for PRAD (Gleason grade group).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.990 (0.971–1.000)
0.986 (0.968–1.000)
1.000 (1.000–1.000)
0.914 (0.810–1.000)
0.992 (0.981–1.000)
GigaPath
0.993 (0.982–1.000)
0.972 (0.949–0.991)
1.000 (1.000–1.000)
0.829 (0.690–0.944)
0.984 (0.969–0.995)
H-optimus-0
0.992 (0.984–1.000)
0.963 (0.935–0.986)
0.995 (0.983–1.000)
0.800 (0.667–0.929)
0.978 (0.962–0.992)
H-optimus-1
0.996 (0.990–1.000)
0.968 (0.945–0.991)
0.995 (0.983–1.000)
0.829 (0.700–0.946)
0.981 (0.966–0.994)
Phikon-v2
0.996 (0.991–1.000)
0.963 (0.935–0.986)
0.989 (0.972–1.000)
0.829 (0.697–0.943)
0.978 (0.961–0.992)
UNI
0.997 (0.994–1.000)
0.972 (0.949–0.991)
1.000 (1.000–1.000)
0.829 (0.700–0.946)
0.984 (0.970–0.995)
Appendix
Extended Data Table 77: Test-set metrics for PRAD (Benign vs. malignant).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
GigaPath
0.999 (0.998–1.000)
0.982 (0.965–0.996)
0.995 (0.984–1.000)
0.919 (0.818–1.000)
0.990 (0.978–0.997)
H-optimus-0
0.999 (0.998–1.000)
0.982 (0.965–0.996)
0.995 (0.984–1.000)
0.919 (0.818–1.000)
0.989 (0.978–0.997)
H-optimus-1
0.999 (0.998–1.000)
0.991 (0.978–1.000)
0.989 (0.973–1.000)
1.000 (1.000–1.000)
0.995 (0.986–1.000)
Phikon-v2
1.000 (0.999–1.000)
0.987 (0.969–1.000)
1.000 (1.000–1.000)
0.919 (0.828–1.000)
0.992 (0.982–1.000)
UNI
1.000 (0.999–1.000)
0.987 (0.969–1.000)
0.995 (0.984–1.000)
0.946 (0.862–1.000)
0.992 (0.982–1.000)
Appendix
Extended Data Table 78: Test-set metrics for STAD (Benign vs. malignant).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
1.000 (1.000–1.000)
0.995 (0.984–1.000)
0.994 (0.981–1.000)
1.000 (1.000–1.000)
0.997 (0.990–1.000)
GigaPath
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
H-optimus-0
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
H-optimus-1
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
Phikon-v2
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
UNI
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
1.000 (1.000–1.000)
Appendix
Extended Data Table 79: Test-set metrics for THCA (Benign vs. malignant).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.985 (0.971–0.999)
0.970 (0.955–0.983)
0.975 (0.960–0.988)
0.946 (0.893–0.989)
0.982 (0.973–0.990)
GigaPath
0.981 (0.971–0.992)
0.843 (0.812–0.875)
0.998 (0.993–1.000)
0.098 (0.040–0.168)
0.913 (0.894–0.932)
H-optimus-0
0.986 (0.971–1.000)
0.966 (0.951–0.979)
0.964 (0.946–0.980)
0.978 (0.941–1.000)
0.979 (0.969–0.988)
H-optimus-1
0.986 (0.968–1.000)
0.961 (0.942–0.976)
0.955 (0.935–0.973)
0.989 (0.963–1.000)
0.976 (0.965–0.985)
Phikon-v2
0.961 (0.940–0.982)
0.847 (0.815–0.877)
0.998 (0.993–1.000)
0.120 (0.055–0.190)
0.915 (0.896–0.933)
UNI
0.970 (0.953–0.986)
0.920 (0.897–0.942)
0.968 (0.952–0.983)
0.685 (0.594–0.779)
0.952 (0.937–0.966)
Appendix
Extended Data Table 80: External-set metrics for UCEC/UCS (Benign vs. malignant).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.862 (0.840–0.884)
0.762 (0.737–0.788)
0.635 (0.579–0.691)
0.808 (0.783–0.836)
0.588 (0.540–0.632)
GigaPath
0.872 (0.851–0.894)
0.840 (0.820–0.862)
0.444 (0.386–0.502)
0.985 (0.976–0.993)
0.598 (0.541–0.653)
H-optimus-0
0.884 (0.864–0.904)
0.830 (0.808–0.852)
0.430 (0.373–0.488)
0.975 (0.964–0.985)
0.574 (0.514–0.629)
H-optimus-1
0.875 (0.854–0.896)
0.825 (0.802–0.847)
0.580 (0.522–0.638)
0.914 (0.894–0.932)
0.639 (0.590–0.685)
Phikon-v2
0.841 (0.817–0.865)
0.683 (0.655–0.711)
0.805 (0.758–0.850)
0.638 (0.607–0.672)
0.576 (0.533–0.615)
UNI
0.878 (0.857–0.898)
0.823 (0.801–0.847)
0.382 (0.325–0.439)
0.984 (0.975–0.992)
0.536 (0.474–0.593)
Appendix
Extended Data Table 81: External-set metrics for LUAD/LUSC (Histological subtype).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
GigaPath
0.989 (0.982–0.997)
0.975 (0.965–0.985)
0.982 (0.971–0.991)
0.965 (0.946–0.981)
0.980 (0.972–0.987)
H-optimus-0
0.993 (0.987–0.998)
0.983 (0.974–0.990)
0.988 (0.979–0.995)
0.974 (0.958–0.988)
0.986 (0.979–0.992)
H-optimus-1
0.992 (0.986–0.998)
0.985 (0.978–0.992)
0.990 (0.981–0.997)
0.979 (0.965–0.991)
0.988 (0.982–0.993)
Phikon-v2
0.989 (0.981–0.997)
0.967 (0.956–0.978)
0.990 (0.981–0.997)
0.932 (0.907–0.956)
0.973 (0.964–0.982)
UNI
0.989 (0.982–0.997)
0.976 (0.967–0.985)
0.985 (0.976–0.994)
0.963 (0.944–0.980)
0.981 (0.973–0.988)
Virchow
0.992 (0.985–0.998)
0.983 (0.974–0.990)
0.991 (0.983–0.997)
0.970 (0.952–0.984)
0.986 (0.979–0.992)
Appendix
Extended Data Table 82: External-set metrics for LUNG (Benign vs. malignant).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.964 (0.947–0.980)
0.920 (0.893–0.943)
0.500 (0.340–0.658)
0.960 (0.940–0.978)
0.519 (0.369–0.651)
GigaPath
0.935 (0.912–0.959)
0.883 (0.850–0.911)
0.725 (0.578–0.857)
0.898 (0.867–0.927)
0.518 (0.393–0.625)
H-optimus-0
0.961 (0.944–0.979)
0.937 (0.913–0.959)
0.575 (0.415–0.729)
0.971 (0.955–0.986)
0.613 (0.464–0.733)
H-optimus-1
0.960 (0.942–0.978)
0.928 (0.902–0.950)
0.425 (0.273–0.585)
0.976 (0.961–0.990)
0.507 (0.346–0.648)
Phikon-v2
0.956 (0.936–0.976)
0.926 (0.900–0.948)
0.700 (0.559–0.833)
0.948 (0.925–0.967)
0.622 (0.494–0.727)
UNI
0.965 (0.949–0.981)
0.917 (0.891–0.941)
0.850 (0.733–0.952)
0.924 (0.898–0.948)
0.642 (0.524–0.737)
Appendix
Extended Data Table 83: External-set metrics for KIRC/KIRP/KICH (Histological subtype).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
FOCUS
0.616 (0.572–0.660)
0.510 (0.462–0.560)
0.277 (0.254–0.305)
0.771 (0.757–0.787)
0.437 (0.387–0.495)
GigaPath
0.541 (0.498–0.587)
0.500 (0.455–0.548)
0.260 (0.243–0.277)
0.762 (0.750–0.774)
0.403 (0.347–0.460)
H-optimus-0
0.598 (0.556–0.643)
0.507 (0.457–0.555)
0.274 (0.251–0.299)
0.765 (0.751–0.780)
0.428 (0.372–0.484)
H-optimus-1
0.614 (0.571–0.658)
0.502 (0.452–0.552)
0.268 (0.248–0.289)
0.767 (0.753–0.782)
0.429 (0.374–0.484)
Phikon-v2
0.660 (0.621–0.700)
0.533 (0.486–0.581)
0.304 (0.274–0.332)
0.787 (0.770–0.803)
0.479 (0.427–0.531)
UNI
0.613 (0.571–0.658)
0.495 (0.448–0.543)
0.269 (0.246–0.295)
0.770 (0.755–0.785)
0.429 (0.374–0.483)
Appendix
Extended Data Table 84: External-set metrics for KIRC/KIRP/KICH (Tumor grading).
Model
AUC
Accuracy
Sensitivity
Specificity
F1 score
H-optimus-0
0.999 (0.997–1.000)
0.975 (0.963–0.985)
0.978 (0.964–0.990)
0.968 (0.943–0.990)
0.981 (0.972–0.989)
H-optimus-1
0.999 (0.999–1.000)
0.979 (0.968–0.990)
0.972 (0.955–0.986)
0.995 (0.984–1.000)
0.985 (0.975–0.992)
Phikon-v2
0.997 (0.995–0.999)
0.966 (0.953–0.978)
0.980 (0.967–0.992)
0.936 (0.900–0.966)
0.975 (0.965–0.984)
UNI
0.999 (0.999–1.000)
0.984 (0.974–0.993)
0.993 (0.985–1.000)
0.964 (0.937–0.986)
0.988 (0.981–0.995)
Virchow2
1.000 (0.999–1.000)
0.991 (0.984–0.997)
0.996 (0.989–1.000)
0.982 (0.961–0.996)
0.993 (0.988–0.998)
Virchow
0.999 (0.998–1.000)
0.978 (0.966–0.988)
0.974 (0.958–0.987)
0.986 (0.970–1.000)
0.984 (0.975–0.991)
Appendix
Extended Data Table 85: External-set metrics for KIRC/KIRP/KICH (Benign vs. malignant).
Model
accuracy
Macro-F1 (%)
Balanced accuracy
FOCUS
73.04 ± 0.96
72.82 ± 0.99
73.04 ± 0.96
UNI
70.84 ± 0.86
70.61 ± 0.86
70.84 ± 0.86
Virchow2
70.68 ± 0.83
70.38 ± 0.85
70.68 ± 0.83
Virchow
70.23 ± 0.41
70.00 ± 0.42
70.23 ± 0.41
H-Optimus-0
70.06 ± 0.98
69.81 ± 0.98
70.06 ± 0.98
GigaPath
69.99 ± 0.63
69.80 ± 0.62
69.99 ± 0.63
Appendix
Extended Data Table 86: Five-fold patch-level classification performance.
Task
AUC, 224 px
AUC, 448 px
AUC, 896 px
Δ AUC, 448–224 px
Δ AUC, 896–224 px
T1
0.964
0.971
0.971
+0.007
+0.007
T2
0.995
0.997
0.993
+0.001
-0.003
T3
0.992
0.988
0.983
-0.004
-0.009
T4
0.987
0.985
0.981
-0.002
-0.006
T5
0.898
0.917
0.935
+0.018
+0.037
T6
0.992
0.997
0.987
+0.005
-0.005
Appendix
Extended Data Table 87: Task-level patch-size robustness of FOCUS.
Contrast (Full FOCUS minus counterpart)
Tasks with Δ AUC >0
Mean Δ AUC
Bootstrap 95% CI
Wilcoxon P
FAMV
17/23
+0.0012
+0.0000 to +0.0022
0.0327
HCDC
22/23
+0.0125
+0.0085 to +0.0173
<0.0001
Appendix
Extended Data Table 88: Matched-task data-pipeline ablation results for FOCUS.
Pathological assessment guides lung cancer diagnosis, treatment selection, and prognostic evaluation, yet current CPath approaches rely on task-specific models for isolated objectives. Although pan-cancer foundation models offer versatility, they lack subspecialty-level depth and have not been evaluated across clinical workflows or prospectively validated in real-world settings. We introduce PulmoFoundation, a multi-center, prospectively validated, randomized controlled trial (RCT)-evaluated foundation model for comprehensive lung pathology assessment across pre-operative, intra-operative, and post-operative care. Built upon Virchow2 via subspecialty-specific pretraining using ~40,000 diagnostic H&E-stained whole-slide images (WSIs), PulmoFoundation was systematically evaluated on ~26,000 WSIs across 32 clinically relevant tasks. In addition to accurately predicting molecular markers and patient survival, our model achieves clinical-grade performance in core diagnostic tasks across biopsy, frozen section, and surgical resection slides. In a registered prospective study of 1,357 patients across 11 diagnostic tasks, our model achieved an average AUC of 92.3%. Using pre-specified triage thresholds, PulmoFoundation could reduce additional second-review burden for 68.8% of biopsies and 83.0% of frozen sections, and defer 44.5% of IHC stain orders, with PPVs of 1.0, 0.991, and 0.966. Beyond prospective validation, we conducted a crossover RCT with eight pathologists, in which AI assistance improved diagnostic accuracy across 4,928 case-reader pairs (91.7% w/ AI vs. 83.8% w/o AI). AI assistance also reduced median diagnostic time by 19.6%, increased diagnostic confidence by 8.7%, and improved inter-rater agreement from moderate (kappa = 0.56) to substantial (kappa = 0.76). Together, these evaluations support PulmoFoundation as a clinically validated decision-support system for lung pathology.
Zhengrui Guo, Zhengyu Zhang, Jiabo Ma +23
Department of Computer Science and Engineering, Hong Kong University of Science and Technology, Hong Kong, China · Department of Pathology, Nanfang Hospital, Southern Medical University, Guangzhou, China · Department of Pathology, School of Basic Medical Sciences, Southern Medical University, Guangzhou, China +15
Breast core needle biopsy (CNB) is central to breast cancer diagnosis yet remains challenging because limited tissue sampling, lesion heterogeneity, and subtle morphologic overlap can obscure subtype distinctions. We developed CorePath, a breast-specialized multimodal pathology foundation model fine-tuned from PRISM using 7901 paired CNB whole-slide images and diagnostic reports from two centers. Evaluated across six CNB cohorts and two public breast pathology benchmarks without task-specific retraining, CorePath consistently outperformed PRISM across cancer detection, invasion assessment, and histological subtyping. It achieved weighted area under the receiver operating characteristic curves (AUCs) of 0.9526-0.9735 for five-class CNB histological subtyping across private centers. On public benchmarks, CorePath outperformed leading pathology foundation models, achieving the highest weighted AUCs of 0.7780 for BCNB invasive carcinoma subtyping, 0.8178 for BRACS lesion stratification, and 0.8252 for BRACS fine-grained classification. In report generation, CorePath reduced the overall non-breast hallucinations from 30.1% to 2.8%, demonstrating improved domain fidelity after breast-specific adaptation. CorePath-CRG further combined conformal filtering of subtype and binary cancer status predictions with Learn-Then-Test-based threshold calibration to support selective narrative release, diagnostic fallback, and deferral. CorePath-CRG achieved zero non-breast hallucinations among released outputs and showed the strongest overall performance in pathologist-validated LLM-based Evaluation Scores and quantitative report-generation metrics across most centers. These results demonstrate that domain-specialized foundation models with statistical risk control offer a promising approach for accurate breast CNB diagnosis and reliable report generation.
Ting Yin, Danning Li, Chen Shu +17
Department of Pathology, West China Hospital, Sichuan University, Chengdu, China · Institute of Clinical Pathology, West China Hospital, Sichuan University, Chengdu, China · The Hong Kong University of Science and Technology (Guangzhou) +7
Pathologists integrate morphology across magnifications and across the slides of a patient case, whereas pathology foundation models encode thousands of tiles from single slides and aggregate their features. Here we present WILSON, a vision--language foundation model that represents whole-slide images and multi-slide cases as single multi-magnification composite images, trained on approximately 189k slides from Mayo Clinic spanning 42 organs and 829 diagnostic entities using pathology reports as supervision. Without task-specific training, WILSON exceeded a dedicated case-level model on all internal cohorts (macro-F1 0.52 versus 0.38) and matched slide-level models up to 9.4 times larger at 272- to 2,155-fold lower compute. End-to-end fine-tuning on 508 triple-negative breast cancer cases improved histologic subtyping and stromal tumor-infiltrating lymphocyte grading by 0.16 and 0.11 macro-F1. WILSON retrieved matching diagnostic text at 75.6% recall@1 (PRISM, 58.1%) and generated captions closer to report-derived references than PRISM and PRISM2 on the internal cohort and on most external comparisons. Composite images thus offer a compact, clinically aligned computational unit for pathology.
Saghir Alfasly, Wataru Uegami, Sobhan Hemati +13
KIMIA Lab, Department of AI & Informatics, Mayo Clinic, Rochester, MN, USA. · Department of Laboratory Medicine and Pathology, Mayo Clinic, Rochester, MN, USA. · Department of Quantitative Health Sciences, Mayo Clinic, Rochester, MN, USA. +3