CARE: A Molecular-Guided Foundation Model with Adaptive Region Modeling for Whole Slide Image Analysis

Di Zhang; Zhangpeng Gong; Xiaobo Pang; Jiashuai Liu; Junbo Lu; Hao Cui; Jiusong Ge; Zhi Zeng

cs.CVFeb 25, 2026

CARE: A Molecular-Guided Foundation Model with Adaptive Region Modeling for Whole Slide Image Analysis

Authors: Di Zhang, Zhangpeng Gong, Xiaobo Pang, Jiashuai Liu, Junbo Lu, Hao Cui, Jiusong Ge, Zhi Zeng, +9 more

Organizations: Xi’an Jiaotong University · University of Cambridge · KingMed · BGI Research · A⋆STAR

Abstract

Foundation models have recently achieved impressive success in computational pathology, demonstrating strong generalization across diverse histopathology tasks. However, existing models overlook the heterogeneous and non-uniform organization of pathological regions of interest (ROIs) because they rely on natural image backbones not tailored for tissue morphology. Consequently, they often fail to capture the coherent tissue architecture beyond isolated patches, limiting interpretability and clinical relevance. To address these challenges, we present Cross-modal Adaptive Region Encoder (CARE), a foundation model for pathology that automatically partitions WSIs into several morphologically relevant regions. Specifically, CARE employs a two-stage pretraining strategy: (1) a self-supervised unimodal pretraining stage that learns morphological representations from 34,277 whole-slide images (WSIs) without segmentation annotations, and (2) a cross-modal alignment stage that leverages RNA and protein profiles to refine the construction and representation of adaptive regions. This molecular guidance enables CARE to identify biologically relevant patterns and generate irregular yet coherent tissue regions, selecting the most representative area as ROI. CARE supports a broad range of pathology-related tasks, using either the ROI feature or the slide-level feature obtained by aggregating adaptive regions. Based on only one-tenth of the pretraining data typically used by mainstream foundation models, CARE achieves superior average performance across 33 downstream benchmarks, including morphological classification, molecular prediction, and survival analysis, and outperforms other foundation model baselines overall.

Figures & tables

Figure 1: CARE vs . conventional CPath chunking on WSIs. Panels (a) and (b) depict patch chunks and regular region chunks, respectively. They both impose artificial grids that split tissue, making them inefficient and semantically weak. Panel (c) shows our adaptive region chunks (CARE), which behave like word-level tokens, respect tissue boundaries, and capture meaningful texture, morphology, and cell layout.

Figure 2: CARE architecture and training pipeline. (a) CARE framework with iBOT-style self-supervised pretraining using a teacher–student setup. CARE comprises three modules: the Adaptive Region Generator (ARG), adaptive region self-attention (ARSA), and Semantic and Prior Fusion (SPF). ARG partitions WSIs into morphologically coherent regions. ARSA operates within each adaptive region to derive region level features. SPF aggregates these features into a WSI embedding. (b) Cross-modal pretraining pipeline. RNA and protein encoders produce molecular embeddings, which are aligned to CARE’s WSI embedding via an InfoNCE loss.

Figure 3: (a) Adaptive Region Generator. Based on soft inclusion, each patch retains only its top-3 candidate subregions and masks out the rest. Cosine similarity is then computed to the unmasked candidates, and the patch is assigned to the highest-scoring subregion, yielding an adaptive repartition of patches. (b) Semantic and Prior Fusion. A lightweight module that aggregates adaptive region features into a slide-level embedding.

Task	Head	Mean-pool	CHIEF	PRISM	GigaPath	TANGLE	FEATHER	TITAN	CARE
Morph. Class.	LR	81.2 $\pm$ 0.1	80.2 $\pm$ 0.1	75.6 $\pm$ 0.3	78.3 $\pm$ 0.1	80.9 $\pm$ 0.1	81.1 $\pm$ 0.1	85.4 $\pm$ 0.1	85.7 $\pm$ 0.0
	$k$ NN	73.5 $\pm$ 0.1	71.4 $\pm$ 0.3	68.9 $\pm$ 0.2	67.8 $\pm$ 0.1	75.2 $\pm$ 0.1	72.2 $\pm$ 0.1	83.1 $\pm$ 0.1	82.3 $\pm$ 0.1
	FT	82.6 $\pm$ 0.1	78.0 $\pm$ 0.1	81.4 $\pm$ 0.1	84.9 $\pm$ 0.1	86.1 $\pm$ 0.1	86.7 $\pm$ 0.0	87.7 $\pm$ 0.1	87.0 $\pm$ 0.1
Molecular Class.	LR	67.1 $\pm$ 0.1	66.3 $\pm$ 0.1	69.1 $\pm$ 0.1	66.6 $\pm$ 0.2	69.1 $\pm$ 0.1	67.6 $\pm$ 0.1	67.8 $\pm$ 0.2	69.5 $\pm$ 0.1
	$k$ NN	64.5 $\pm$ 0.1	62.7 $\pm$ 0.1	65.2 $\pm$ 0.1	62.0 $\pm$ 0.1	65.9 $\pm$ 0.1	64.6 $\pm$ 0.1	66.9 $\pm$ 0.1	68.0 $\pm$ 0.1
	FT	70.9 $\pm$ 0.1	73.4 $\pm$ 0.2	73.2 $\pm$ 0.1	73.5 $\pm$ 0.3	74.2 $\pm$ 0.2	75.6 $\pm$ 0.1	71.7 $\pm$ 0.1	74.7 $\pm$ 0.2

Table 1: Average ACC (or C-index) by task category across the 33-task benchmark. The best score is in bold, and the second-best is underlined. “Morph. Class.” denotes the average performance on the morphological classification task. “Molecular Class.” reports the mean result for molecular tasks. “Molecular

\text{Class.}_{V}

” denotes the average result on no-validation subset of molecular tasks. “Sur. Anal.” denotes to the average performance on the survival prediction task. Detailed results for the 33 benchmarks are provided in the appendix.

Figure 4: AUROC (or F1) results on 33 downstream tasks under a logistic regression (or a linear layer for survival analysis) setting. The best-performing results are outlined with black boxes.

Figure 5: Box plots of experimental results. Each box plot aggregates results from all morphology-classification and molecular-classification tasks. The horizontal line inside each box indicates the median for that model across tasks. The green circle denotes the mean.

Figure 6: Heatmap visualizations across different foundation models. The WSIs are sourced from the CPTAC-HNSCC dataset. All saliency maps were post-processed with Gaussian blur to smooth region boundaries. Warmer colors (red) indicate stronger model attention.

Method	Slide-feature		ROI-feature
Metric	ACC	AUC	ACC	AUC
CCRCC-BAP1	58.7	65.8	61.0	69.1
HNSCC-CASP8	51.7	58.7	53.7	61.5
Cross-LUNG-subtype	52.5	77.0	56.4	81.2

Table 2: Results for ROI feature classification. WSI-feature denotes linear-probing using CARE’s WSI embedding. ROI-feature denotes linear-probing using ROI embeddings of CARE.

Method	EBRAINS-fine		BCNB-HER2
Metric	ACC	F1	ACC	AUC
Not adaptive region	71.6	73.4	61.4	75.0
Not $g^{\mathrm{CLS}}_{i}$ feature	71.5	71.9	58.9	73.8
Not $g^{Q}_{i}$ feature	71.2	71.7	61.9	74.5
Not $f_{i}$ feature	70.1	69.9	61.6	75.6
Not $f^{R}_{i}$ feature	70.8	71.2	59.6	75.4

Table 3: Ablation study on the ARG. Not adaptive region denotes using fixed subregions in place of adaptive regions. All results are reported under the random initialization setting.

	EBRAINS-fine		BCNB-HER2
Metric	ACC	F1	ACC	AUC
Window size $k=6$	71.2	73.4	59.4	73.3
Window size $k=8$	72.5	74.2	64.8	75.4
Window size $k=10$	71.9	73.9	61.9	75.1
$\lambda_{\mathrm{RSL}}=0.01$	72.1	73.4	60.2	74.7
$\lambda_{\mathrm{RSL}}=0.1$	72.5	74.2	64.8	75.4

Table 4: Ablation results for CARE hyperparameters. All experiments are reported under the random initialization setting.

Method	EBRAINS-fine		BCNB-HER2
Metric	ACC	F1	ACC	AUC
CARE-ibot (less)	71.5	76.5	56.9	76.8
CARE-ibot	72.2	77.4	60.3	77.0
CARE-RNA	73.7	78.5	59.7	76.9
CARE	74.0	78.7	59.8	77.2

Table 5: Ablation results across pretraining stages. All results are reported under the linear probing setting using logistic regression. CARE-iBOT (less) indicates the version of the model trained via iBOT using less data (24,000 WSIs).

Hyperparameter	Value
Regional self-attention depth	2
Regional self-attention heads	8
Regional cross-attention heads	8
Adaptive region self-attention depth	5
Adaptive region self-attention heads	8
Window size	8

Table 6: Hyperparameter Selection for CARE.

Hyperparameter	Value
Global crop patch ratio (% per sub-WSI)	90%
Global crop number	2
Local crop patch ratio (% per sub-WSI)	18%
Local crop number	6
Prediction mask ratio	0.3
Prediction mask variance	0.15

Table 7: Hyperparameter Selection during Unimodal Self-Supervised Pretraining. We use 8×80 GB A100 GPUs for the unimodal pretraining stage.

Hyperparameter	Value
Gene number	3999
Dropout	0.2
WSI encoder learning rate	1e-5
Minimum WSI encoder learning rate	1e-6
RNA encoder learning rate	5e-5
Minimum RNA encoder learning rate	1e-5

Table 8: Hyperparameter Selection during molecule-slide cross-modal contrastive pretraining. We use 8×80 GB A100 GPUs for the cross-modal pretraining stage.

Models	Patch encoder	Slide aggregator	Output size	FLOPs/G	Para/M	Dataset size
CHIEF	Ctranspath	Variant of ABMIL	$1\times 768$	0.26	1.05	60,530
PRISM	Virchow	Perceiver	$1\times 1280$	359.36	89.21	587,196
GigaPath	GigaPath	LongNet	$1\times 768$	31.80	86.33	171,189
TANGLE	UNI	ABMIL	$1\times 512$	1.57	5.00	$24{,}000^{*}$
FEATHER	CONCH v1.5	ABMIL	$1\times 512$	0.31	0.79	$24{,}000^{*}$
TITAN	CONCH v1.5	ViT	$1\times 768$	19.21	47.36	$335,645(423{,}122/182{,}862)$

Table 9: Architectural and pretraining statistics of compared slide-level foundation models. We report only the parameters of the vision-inference module. The superscript * indicates that the model was pre-trained on the TCGA pan-cancer dataset and did not report specific numbers.

Task	Metric	Mean-pool	CHIEF	PRISM	GigaPath	TANGLE	FEATHER	TITAN	CARE
Task 1	ACC	77.8 $\pm$ 0.1	73.0 $\pm$ 0.1	72.2 $\pm$ 0.2	75.8 $\pm$ 0.1	76.4 $\pm$ 0.1	82.3 $\pm$ 0.1	87.1 $\pm$ 0.0	85.1 $\pm$ 0.1
Task 1	F1	87.1 $\pm$ 0.0	83.9 $\pm$ 0.0	82.9 $\pm$ 0.0	87.0 $\pm$ 0.0	86.7 $\pm$ 0.0	90.0 $\pm$ 0.0	91.7 $\pm$ 0.0	91.2 $\pm$ 0.0
Task 2	ACC	65.8 $\pm$ 0.0	60.6 $\pm$ 0.1	59.5 $\pm$ 0.0	64.7 $\pm$ 0.1	64.5 $\pm$ 0.0	68.2 $\pm$ 0.0	74.8 $\pm$ 0.0	74.0 $\pm$ 0.1
Task 2	F1	71.4 $\pm$ 0.0	68.4 $\pm$ 0.1	65.8 $\pm$ 0.0	71.6 $\pm$ 0.1	71.2 $\pm$ 0.0	73.1 $\pm$ 0.0	78.8 $\pm$ 0.0	78.7 $\pm$ 0.0
Task 3	ACC	90.0 $\pm$ 0.0	91.4 $\pm$ 0.0	91.9 $\pm$ 0.0	91.1 $\pm$ 0.0	91.0 $\pm$ 0.0	91.0 $\pm$ 0.0	92.8 $\pm$ 0.0	92.6 $\pm$ 0.0
Task 3	F1	90.6 $\pm$ 0.0	91.4 $\pm$ 0.0	92.2 $\pm$ 0.0	91.3 $\pm$ 0.0	91.3 $\pm$ 0.0	91.3 $\pm$ 0.0	92.9 $\pm$ 0.0	92.7 $\pm$ 0.0

Table 10: Results under the linear probing (logistic regression) setting for morphologic and molecular classification tasks. The best score is in bold, and the second-best is underlined.

Task	Metric	Mean-pool	CHIEF	PRISM	GigaPath	TANGLE	FEATHER	TITAN	CARE
Task 15	ACC	59.1 $\pm$ 0.3	56.2 $\pm$ 0.2	51.9 $\pm$ 0.2	57.1 $\pm$ 0.2	58.1 $\pm$ 0.2	59.8 $\pm$ 0.2	57.9 $\pm$ 0.2	57.2 $\pm$ 0.3
Task 15	AUC	61.7 $\pm$ 0.4	63.3 $\pm$ 0.2	53.9 $\pm$ 0.3	62.8 $\pm$ 0.2	64.8 $\pm$ 0.3	62.5 $\pm$ 0.3	62.3 $\pm$ 0.4	61.3 $\pm$ 0.2
Task 16	ACC	75.8 $\pm$ 0.7	72.8 $\pm$ 0.8	67.9 $\pm$ 0.6	71.8 $\pm$ 1.0	75.0 $\pm$ 0.8	62.7 $\pm$ 1.2	72.9 $\pm$ 0.8	75.3 $\pm$ 0.8
Task 16	AUC	83.5 $\pm$ 0.9	81.4 $\pm$ 1.1	78.9 $\pm$ 0.6	81.7 $\pm$ 0.9	83.6 $\pm$ 0.9	70.3 $\pm$ 1.4	79.9 $\pm$ 0.9	81.2 $\pm$ 0.8
Task 17	ACC	63.1 $\pm$ 1.5	64.1 $\pm$ 1.4	54.8 $\pm$ 1.2	62.0 $\pm$ 0.9	55.1 $\pm$ 0.7	59.6 $\pm$ 1.2	62.8 $\pm$ 1.0	64.5 $\pm$ 0.8
Task 17	AUC	69.1 $\pm$ 1.5	70.6 $\pm$ 1.8	59.3 $\pm$ 1.5	72.2 $\pm$ 1.4	67.0 $\pm$ 1.2	63.9 $\pm$ 1.2	68.1 $\pm$ 1.1	66.6 $\pm$ 1.1

Table 11: Results on linear probing (logistic regression) setting for molecular classification tasks without a validation set. The best score is in bold, and the second-best is underlined.

Task	Metric	Mean-pool	CHIEF	PRISM	GigaPath	TANGLE	FEATHER	TITAN	CARE
Task 1	ACC	71.4 $\pm$ 0.0	55.8 $\pm$ 0.1	57.8 $\pm$ 0.1	58.7 $\pm$ 0.1	66.1 $\pm$ 0.2	70.8 $\pm$ 0.1	83.5 $\pm$ 0.1	80.9 $\pm$ 0.0
Task 1	F1	83.8 $\pm$ 0.0	72.4 $\pm$ 0.0	73.1 $\pm$ 0.0	75.4 $\pm$ 0.0	80.7 $\pm$ 0.0	84.0 $\pm$ 0.0	90.7 $\pm$ 0.0	89.4 $\pm$ 0.0
Task 2	ACC	56.6 $\pm$ 0.0	44.8 $\pm$ 0.1	48.0 $\pm$ 0.0	46.4 $\pm$ 0.0	54.1 $\pm$ 0.0	56.9 $\pm$ 0.0	70.5 $\pm$ 0.0	66.2 $\pm$ 0.0
Task 2	F1	63.3 $\pm$ 0.0	53.3 $\pm$ 0.1	54.5 $\pm$ 0.0	55.4 $\pm$ 0.0	62.8 $\pm$ 0.0	65.2 $\pm$ 0.0	75.8 $\pm$ 0.0	72.0 $\pm$ 0.0
Task 3	ACC	84.5 $\pm$ 0.0	90.3 $\pm$ 0.0	88.6 $\pm$ 0.0	84.5 $\pm$ 0.0	87.8 $\pm$ 0.0	81.1 $\pm$ 0.0	90.5 $\pm$ 0.0	89.7 $\pm$ 0.0
Task 3	F1	86.7 $\pm$ 0.0	90.3 $\pm$ 0.0	89.4 $\pm$ 0.0	86.6 $\pm$ 0.0	89.1 $\pm$ 0.0	85.6 $\pm$ 0.0	91.3 $\pm$ 0.0	90.4 $\pm$ 0.0

Table 12: Results on linear probing (

k

NN parameter) setting for morphologic and molecular classification tasks. The best score is in bold, and the second-best is underlined.

Task	Metric	Mean-pool	CHIEF	PRISM	GigaPath	TANGLE	FEATHER	TITAN	CARE
Task 15	ACC	55.2 $\pm$ 0.1	54.4 $\pm$ 0.2	50.9 $\pm$ 0.1	54.7 $\pm$ 0.1	53.3 $\pm$ 0.2	53.2 $\pm$ 0.1	55.3 $\pm$ 0.1	54.3 $\pm$ 0.1
Task 15	AUC	59.4 $\pm$ 0.3	56.1 $\pm$ 0.3	51.2 $\pm$ 0.4	56.7 $\pm$ 0.3	57.5 $\pm$ 0.4	59.2 $\pm$ 0.4	61.3 $\pm$ 0.2	61.3 $\pm$ 0.3
Task 16	ACC	55.9 $\pm$ 0.4	64.1 $\pm$ 0.6	57.5 $\pm$ 0.3	56.2 $\pm$ 0.5	66.8 $\pm$ 0.9	54.5 $\pm$ 0.2	65.0 $\pm$ 0.7	65.5 $\pm$ 0.8
Task 16	AUC	67.6 $\pm$ 1.3	78.5 $\pm$ 1.0	65.0 $\pm$ 0.7	67.1 $\pm$ 1.8	74.8 $\pm$ 1.4	59.5 $\pm$ 1.0	74.5 $\pm$ 1.4	76.9 $\pm$ 1.0
Task 17	ACC	55.1 $\pm$ 0.7	61.1 $\pm$ 0.8	54.5 $\pm$ 0.6	56.0 $\pm$ 0.6	56.1 $\pm$ 0.6	54.0 $\pm$ 0.4	59.4 $\pm$ 0.7	59.0 $\pm$ 0.8
Task 17	AUC	65.5 $\pm$ 1.0	67.5 $\pm$ 1.1	67.5 $\pm$ 1.4	63.8 $\pm$ 1.1	63.4 $\pm$ 1.2	62.3 $\pm$ 1.9	66.9 $\pm$ 1.1	71.4 $\pm$ 1.0

Table 13: Results on linear probing (

k

NN parameter) setting for molecular classification tasks without a validation set. The best score is in bold, and the second-best is underlined.

Task	Mean-pool	CHIEF	PRISM	GigaPath	TANGLE	FEATHER	TITAN	CARE
Task 29	48.4 $\pm$ 16.5	53.2 $\pm$ 15.4	38.0 $\pm$ 20.0	55.8 $\pm$ 4.7	46.6 $\pm$ 8.3	56.6 $\pm$ 15.3	40.5 $\pm$ 18.9	63.0 $\pm$ 8.9
Task 30	53.8 $\pm$ 15.8	58.7 $\pm$ 8.7	39.9 $\pm$ 9.9	57.3 $\pm$ 11.1	48.6 $\pm$ 23.6	57.6 $\pm$ 12.4	38.9 $\pm$ 12.6	65.9 $\pm$ 18.0
Task 31	48.4 $\pm$ 13.8	56.3 $\pm$ 9.6	38.0 $\pm$ 21.8	56.9 $\pm$ 16.1	50.8 $\pm$ 9.2	46.4 $\pm$ 16.0	54.6 $\pm$ 20.6	47.5 $\pm$ 11.9
Task 32	42.2 $\pm$ 7.6	50.3 $\pm$ 10.2	48.3 $\pm$ 8.7	46.8 $\pm$ 8.9	52.2 $\pm$ 8.1	61.3 $\pm$ 10.5	43.5 $\pm$ 6.6	56.6 $\pm$ 10.1
Task 33	49.7 $\pm$ 4.8	53.0 $\pm$ 7.3	46.2 $\pm$ 10.2	62.0 $\pm$ 7.5	50.8 $\pm$ 6.0	47.4 $\pm$ 6.2	58.5 $\pm$ 6.5	57.1 $\pm$ 3.1

Table 14: Survival prediction experimental results. All results are obtained under the linear-probing (Linear) setting. The best score is in bold, and the second-best is underlined.

Figure 7: Experimental results of ACC box plots for different models under different experimental settings. Each box plot aggregates results from all morphology-classification and molecular-classification tasks. The horizontal line inside each box indicates the median for that model across tasks. The green circle denotes the mean.

Figure 8: Heatmap visualization of CARE patches within the ROI. The WSIs are sourced from the CPTAC-HNSCC dataset. Warmer colors (red) indicate stronger patch attention.

Task	Metric	Mean-pool	CHIEF	PRISM	GigaPath	TANGLE	FEATHER	TITAN	CARE
Task 1	ACC	82.6 $\pm$ 0.0	70.6 $\pm$ 0.1	78.9 $\pm$ 0.1	85.2 $\pm$ 0.0	84.5 $\pm$ 0.1	88.0 $\pm$ 0.0	88.3 $\pm$ 0.1	88.9 $\pm$ 0.1
Task 1	F1	83.9 $\pm$ 0.0	75.5 $\pm$ 0.1	85.0 $\pm$ 0.1	90.0 $\pm$ 0.0	88.6 $\pm$ 0.0	89.4 $\pm$ 0.0	92.0 $\pm$ 0.0	91.7 $\pm$ 0.0
Task 2	ACC	67.9 $\pm$ 0.1	57.4 $\pm$ 0.2	63.5 $\pm$ 0.1	71.8 $\pm$ 0.0	71.3 $\pm$ 0.0	74.6 $\pm$ 0.1	75.9 $\pm$ 0.0	73.0 $\pm$ 0.1
Task 2	F1	68.7 $\pm$ 0.1	54.4 $\pm$ 0.2	66.2 $\pm$ 0.1	73.8 $\pm$ 0.0	72.8 $\pm$ 0.0	74.9 $\pm$ 0.0	78.3 $\pm$ 0.0	75.1 $\pm$ 0.1
Task 3	ACC	88.5 $\pm$ 0.0	93.1 $\pm$ 0.0	93.4 $\pm$ 0.0	93.9 $\pm$ 0.0	93.5 $\pm$ 0.0	92.8 $\pm$ 0.0	93.2 $\pm$ 0.0	93.2 $\pm$ 0.0
Task 3	F1	87.7 $\pm$ 0.0	92.9 $\pm$ 0.0	93.4 $\pm$ 0.0	93.9 $\pm$ 0.0	93.6 $\pm$ 0.0	92.8 $\pm$ 0.0	92.5 $\pm$ 0.0	92.9 $\pm$ 0.0

Table 15: Results on fine tuning setting. Since Tasks 15 to 33 do not have a validation set, the fine-tuning experiments were only evaluated on Tasks 1 to 14. The best score is in bold, and the second-best is underlined.

Task	Head	Mean-pool	CHIEF	PRISM	GigaPath	TANGLE	FEATHER	TITAN	CARE
Morph. Class.	LR	87.2 $\pm$ 0.0	86.6 $\pm$ 0.0	84.5 $\pm$ 0.0	85.3 $\pm$ 0.0	87.1 $\pm$ 0.0	87.4 $\pm$ 0.1	90.1 $\pm$ 0.0	90.1 $\pm$ 0.0
	kNN	82.2 $\pm$ 0.0	80.3 $\pm$ 0.1	79.8 $\pm$ 0.0	78.0 $\pm$ 0.0	83.6 $\pm$ 0.0	81.8 $\pm$ 0.0	88.1 $\pm$ 0.0	87.3 $\pm$ 0.0
	FT	82.2 $\pm$ 0.1	79.5 $\pm$ 0.1	85.4 $\pm$ 0.2	86.6 $\pm$ 0.2	87.9 $\pm$ 0.1	88.1 $\pm$ 0.1	89.7 $\pm$ 0.1	88.2 $\pm$ 0.3
Molecular Class.	LR	79.8 $\pm$ 0.1	82.5 $\pm$ 0.1	82.7 $\pm$ 0.1	83.1 $\pm$ 0.1	83.5 $\pm$ 0.1	82.0 $\pm$ 0.1	81.1 $\pm$ 0.2	83.6 $\pm$ 0.1
	kNN	72.5 $\pm$ 0.1	73.8 $\pm$ 0.1	75.7 $\pm$ 0.1	71.3 $\pm$ 0.2	75.5 $\pm$ 0.1	74.5 $\pm$ 0.1	76.7 $\pm$ 0.2	79.1 $\pm$ 0.2
	FT	78.3 $\pm$ 0.1	80.3 $\pm$ 0.1	83.8 $\pm$ 0.1	84.7 $\pm$ 0.1	85.4 $\pm$ 0.1	82.2 $\pm$ 0.1	83.3 $\pm$ 0.1	85.5 $\pm$ 0.1

Table 16: Average AUC (or F1) results across task categories. The best score is in bold, and the second-best is underlined.

Explore similar work

Sep 27, 2026

Can Protein-Derived Knowledge Improve Pathology Foundation Models?

Molecularly guided pathology foundation models (PFMs) exploit transcriptomic or proteomic information to enrich whole-slide image (WSI) representations, yet effectively leveraging large standalone molecular corpora remains challenging. First, existing molecular foundation models encode protein sequences or single-cell states, not the patient-level bulk expression profiles paired with WSIs. Second, because cross-modal supervision is restricted to paired WSI-omics samples, knowledge from standalone molecular corpora reaches the pathology encoder only indirectly, creating a paired-support bottleneck. To address these challenges, we propose a three-stage framework that decouples proteomic knowledge acquisition from cross-modal transfer, yielding ProSlide, a slide-level hierarchical pathology foundation model. First, to close the modality gap, we pretrain a Proteomic Foundation Encoder (PFE) on 12,695 sample-level bulk protein profiles using virtual profile generation and expression-space multi-view pretraining. Second, we pretrain ProSlide, a patch-region-slide encoder, to predict protein expression from paired WSI-protein samples. Third, to relax the paired-support bottleneck, we introduce Prot2Path, a cross-modal relational distillation objective. For each paired sample, it aligns the similarity distributions of the WSI and its protein profile over a shared, frozen bank of PFE-encoded paired and standalone profiles. We evaluate ProSlide on 12 downstream tasks across breast, lung, and renal cancers. Despite being pretrained with only 2,229 WSIs and 12,695 sample-level protein profiles, ProSlide achieves the highest mean accuracy and AUC within each cancer group.

Di Zhang, Zhangpeng Gong, Jiashuai Liu +10

School of Computer Science and Technology, Xi’an Jiaotong University · Tsinghua Shenzhen International Graduate School, Tsinghua University · University of Cambridge +4

Pathology Foundation Models Computational Pathology Whole-Slide Image Analysis Molecular Representation Learning Cross-Modal Knowledge Distillation

Ask PDF

Mar 27, 2026

MOOZY: A Patient-First Foundation Model for Computational Pathology

Computational pathology needs whole-slide image (WSI) foundation models that transfer across diverse clinical tasks, yet current approaches remain largely slide-centric, often depend on private data and expensive paired-report supervision, and do not explicitly model relationships among multiple slides from the same patient. We present MOOZY, a patient-first pathology foundation model in which the patient case, not the individual slide, is the core unit of representation. MOOZY explicitly models dependencies across all slides from the same patient via a case transformer during pretraining, combining multi-stage self-supervision with scaled low-cost task supervision. In Stage 1, we pretrain a vision-only slide encoder on 77,134 public slide feature grids using masked self-distillation. In Stage 2, we align these representations with clinical semantics using a case transformer and multi-task supervision over 333 tasks from 56 public datasets, including 205 classification and 128 survival tasks across four endpoints. Across sixteen held-out tasks, MOOZY improves macro weighted F1, balanced accuracy, and macro weighted ROC-AUC relative to PRISM by +4.19%, +7.93%, and +6.95%, respectively. MOOZY is also parameter efficient with 85.77M parameters, 14

\times

smaller than GigaPath. These results suggest that patient-level pretraining yields transferable embeddings, providing a path toward scalable patient-first histopathology foundation models.

Yousef Kotp, Vincent Quoc-Huy Trinh, Christopher Pal +1

CSSE, Concordia University, Montréal, Canada · Mila – Québec AI Institute, Montréal, Canada · CHUM, Université de Montréal, Montréal, Canada +4

Computational Pathology Self-Supervised Pre-Training Pathology Foundation Models Clinical Prediction Whole-Slide Image Analysis

Ask PDF

Sep 2, 2026

Morphology signal in whole slide image foundation models can automatically triage slides

Patient exams in the cancer diagnosis and staging process typically generate several whole slide images (WSIs). One of the initial steps in training models on WSI data is identifying one or a few slides containing tumor or other diagnostic biomarkers necessary for downstream prediction tasks such as estimating recurrence risk or progression-free survival. This step requires tedious manual curation by experienced pathologists. Many published datasets make the artificial assumption of 1 slide per patient. Alternatively, all slides per patient may be used for model training, which may dilute the signal from the few slides containing tumor or other relevant information. In this paper, we present a pipeline to overcome these challenges using publicly available WSI foundation models (FMs). Our evaluations show that ranking WSIs based on predictions from zero-shot classification using WSI FMs accurately identifies slides with the most tumor, indicating that WSI FMs contain sufficient morphology signal to automatically triage slides. We also present a formulation for ranked evaluation to benchmark FM performance in slide triage. We show, on multiple datasets, that tumor slides are identified in the top-2 ranked slides for patients with up to 43 slides.

Ayushi Sinha, Shashank Yadav, Benjamin Holmes +9

Department of Radiology, Mayo Clinic, Rochester MN · Department of Radiation Oncology, Mayo Clinic, Rochester MN · AI Program, Mayo Clinic, Rochester MN +4

Computational Pathology Pathology Foundation Models Oncology Whole-Slide Image Analysis

Ask PDF