Knowledge-structured and evidence-aware semantic self-distillation for glomerular pathology foundation model

Qiming He; Jing Li; Zihan Wang; Tian Guan; Yifei Ma; Zimo Zhao; Yanxia Wang; Hongjing Chen

cs.CVOct 28, 2026

Knowledge-structured and evidence-aware semantic self-distillation for glomerular pathology foundation model

Authors: Qiming He, Jing Li, Zihan Wang, Tian Guan, Yifei Ma, Zimo Zhao, Yanxia Wang, Hongjing Chen,

Organizations: Fuzhou University Affiliated Provincial Hospital, Interdisciplinary Institute for Medical Engineering, Fuzhou University, Fuzhou, China. · Medical Optical Technology R&D Center, Research Institute of Tsinghua, Pearl River Delta, Guangzhou, China. · Department of Pathology, School of Basic Medicine and Xijing Hospital, Fourth Military Medical University, Xi’an, China. · Institute of Biopharmaceutical and Health Engineering, Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China. ·

Abstract

Visual foundation models learn rich pathological morphology but lack explicit concept-level supervision. Vision--language pretraining can provide this supervision but requires costly image--text pairs. Here, we introduce GloSAGE, a semantic self-distillation framework that uses lesion profiles predicted by a pretrained visual teacher to construct structured pathology descriptions. Reliability-guided selection identifies image--text anchors, while Knowledge-Structured Pathology Text Synthesis and Evidence-Guided Negative-pair Construction account for partially observed lesions. Image--text alignment trains projection layers while retaining the frozen visual encoder; an intermediate Semantic Transfer Layer supplies features for downstream tasks. Developed on GloMillion and evaluated across seven glomerulus- and case-level cohorts, GloSAGE supported zero-shot lesion recognition, image--text retrieval, hard-negative discrimination and concept-intervention analyses. It outperformed its source visual model in most lesion-recognition tasks and improved lesion grading, while supporting competitive case-level disease classification, injury grading and baseline renal function estimation. Ablations supported the combined contributions of knowledge-structured text and evidence-guided negatives, and the intermediate transfer layer improved mean lesion-recognition ROC-AUC relative to the complete projection output. A frozen language model conditioned on GloSAGE features generated glomerular descriptions. These findings suggest that teacher-derived lesion profiles can provide annotation-efficient semantic supervision for extending a glomerular visual foundation model.

Figures & tables

Figure 1: Overview of GloSAGE. (A) Development of GloSAGE through semantic self-distillation. A pretrained visual teacher predicts lesion profiles for glomerular images in GloMillion, and KEYS converts these profiles into knowledge-structured captions. Paired images and captions undergo image–text contrastive pretraining through PIRA with EDGE-guided negative construction. STL transforms pretrained visual features for downstream tasks. GloSAGE is evaluated across seven cohorts for semantic capabilities, clinical diagnosis, component contributions and language-related extensions. (B) Framework of KEYS, PIRA and EDGE. Reliability-guided selection identifies over 40,000 trustworthy visual–semantic anchors. KEYS transforms their lesion profiles into structured descriptions incorporating lesion hierarchies, subtypes and severity grades. PIRA aligns images encoded by the frozen visual encoder with descriptions encoded by PubMedBERT through trainable projection heads, while EDGE excludes unsupported or contradicted negative pairs. STL comprises the first visual projection layer and GELU activation and provides the downstream visual representation. (C) Evaluation of semantic enhancement through zero-shot lesion recognition, bidirectional retrieval, hard-negative discrimination and concept intervention. (D) Evaluation of the diagnostic benefits of semantic enhancement at the glomerular and case levels. Glomerular evaluations include the distribution of ROC-AUC across 27 lesion-recognition tasks (left) and performance on 15 tasks across four independent cohorts (middle). Case-level evaluations include kidney disease diagnosis, subtyping, injury grading and biomarker prediction. (E) Ablation results assessing the contributions of individual components. (F) Exploratory evaluation of language-related capabilities.

Figure 2: Semantic assessment of GloSAGE across three histological stains. (A) Experimental setup and performance comparison for zero-shot lesion recognition. The bar chart shows the overall ROC-AUC of each model across all evaluated lesions for the three stains. The three bubble plots illustrate the lesion-wise ROC-AUC rankings of the models for PAS, Masson and PASM, respectively. (B) Experimental setup and performance comparison for bidirectional image–text retrieval, evaluated using R@1, R@5 and R@10. The upper and lower rows of line plots represent image-to-text and text-to-image retrieval, respectively, while the three columns correspond to PAS, Masson and PASM. (C) Ranking a matched lesion description against five hard negatives. Heatmaps show the difference between hard-negative accuracy and the six-candidate chance level (1/6) for individual lesions, while bar charts summarize the mean difference across all evaluated stress-test targets for each stain. (D) Experimental setup and performance comparison for concept intervention. The three horizontal bar charts illustrate the distribution of the best-performing model–lesion combinations across the three stains, while the vertical bar chart compares the overall ROC-AUC across models.

Figure 3: GloSAGE enhances diagnostic performance across multiple levels of glomerular pathology analysis. (A) Performance comparison of glomerular lesion recognition across different foundation models. The upper-left panel shows the mean ROC-AUC across all lesions for each of the three stains. The upper-right, lower-left, and lower-right panels show the relative reduction in ROC-AUC deficit achieved by GloSAGE compared with other models on PAS, Masson, and PASM stains, respectively. (B) Representative visualizations of glomerular lesion recognition by GloSAGE. The three rows correspond to PAS, Masson, and PASM stains, respectively, while the three columns represent mesangial hypercellularity (MH), crescent formation (Cre), and global sclerosis (GS). (C) Performance comparison of glomerular lesion grading. The left panel shows the distributions of ROC-AUC across multiple cross-validation folds for different models, while the right panel illustrates the distribution of ROC-AUC improvements achieved by GloSAGE over GloPath. (D) Representative visualizations of glomerular lesion grading by GloSAGE. The three rows correspond to PAS, Masson, and PASM stains, respectively, and the three columns illustrate different severity grades or morphological subtypes. (E) Performance on GloGS-1 (recognition of GS), GloGS-2 (recognition of GS), GloGrade (injury grading of glomeruli). (F) Performance on GloPro (lesion recognition on PAS).

Figure 4: (A) Case-level performance comparison for disease classification (top row) and injury grading (bottom row) on RenalXJ. (B) Representative examples of the top 12 glomeruli ranked by attention weights in case-level disease classification and injury grading on RenalXJ. (C) Case-level performance comparison for disease classification (top row) and biomarker regression (bottom row) on RenalKPMP. (D) Representative examples of the top 12 glomeruli ranked by attention weights in case-level disease classification and biomarker regression on RenalKPMP.

Figure 5: Ablation of the semantic components of GloSAGE. (A) GloSAGE architecture highlighting KEYS, EDGE and STL. (B) ROC-AUC rankings of the ablation configurations across individual lesions. (C) Mean ROC-AUC for each stain. (D) GloSAGE gains relative to each ablation configuration on lesion-recognition tasks. Rows correspond to PAS, Masson and PASM, respectively; columns correspond to w/o EDGE, w/o KEYS and w/o EDGE and KEYS, respectively. (E) Comparison of STL features with the complete projection-head output (GloSAGE-P). The upper-left panel shows mean ROC-AUC across the three stains; the upper-right, lower-left and lower-right panels show lesion-specific ROC-AUC for PAS, Masson and PASM, respectively.

Figure 6: Image-conditioned generation of glomerular descriptions. (A) A trainable visual-prefix adapter converts features from a frozen visual encoder into eight tokens that condition a frozen language model. Each image is paired with the same fixed descriptive instruction; knowledge-graph-derived descriptions provide the training references. (B) ROUGE-1, ROUGE-2, and ROUGE-L recall scores against structured reference descriptions across PAS, Masson, and PASM. (C) Selected generated descriptions from the GloSAGE-based system alongside GPT-6-Astra outputs for the same images. Panel C is qualitative and is not a quantitative comparison between the two systems.

Appendix figures & tables33 assets

Supplementary material from the paper’s appendix.

Appendix

Figure S1: Glomerular lesion knowledge graph and structured-text construction. The selected hierarchy and node attributes illustrate how KEYS retrieves concept names, definitions and synonyms to assemble pathological descriptions. Representative examples show global sclerosis in PAS, loop necrosis with a cellular crescent in Masson, and no positive findings in PASM. Crescent subtypes are consolidated into a single lesion group during profile construction. The empty-profile description is restricted to the assessed lesion set.

Figure S2: Comparison of models on zero-shot tasks on three stains.

Figure S3: Lesion-wise concept-intervention effect sizes: mean change in image–text similarity on images positive for the original lesion, after replacing its description with a competing concept.

Figure S4: Performance rankings of models across 27 lesion recognition tasks.

Lesion	GloSAGE	GloPath	KEEP	PLIP	CONCH
PAS-MH	0.703	0.432	0.584	0.421	0.439
PAS-PAS(+)	0.828	0.356	0.458	0.425	0.443
PAS-PAS(-)	0.816	0.479	0.478	0.575	0.643
PAS-Crum	0.567	0.674	0.525	0.556	0.571
PAS-ADH	0.513	0.389	0.486	0.455	0.330
PAS-SS	0.749	0.491	0.537	0.508	0.559

Appendix

Table S1: ROC-AUC of zero-shot task on lesion recognition (PAS)

Lesion	GloSAGE	GloPath	KEEP	PLIP	CONCH
MT-MH	0.778	0.384	0.440	0.502	0.416
MT-ADH	0.470	0.513	0.585	0.623	0.390
MT-SS	0.633	0.570	0.693	0.417	0.469
MT-SFN	0.837	0.572	0.670	0.545	0.647
MT-Cre	0.821	0.667	0.643	0.583	0.823
MT-GS	0.596	0.745	0.290	0.232	0.865

Appendix

Table S2: ROC-AUC of zero-shot task on lesion recognition (Masson)

Lesion	GloSAGE	GloPath	KEEP	PLIP	CONCH
PASM-MH	0.776	0.498	0.418	0.513	0.408
PASM-PAS(-)	0.837	0.638	0.463	0.543	0.771
PASM-ADH	0.654	0.454	0.473	0.525	0.378
PASM-SS	0.768	0.484	0.395	0.493	0.392
PASM-SFN	0.827	0.715	0.634	0.509	0.558
PASM-MN	0.670	0.483	0.603	0.571	0.504

Appendix

Table S3: ROC-AUC of zero-shot task on lesion recognition (PASM)

	Image-to-Text			Text-to-Image
Model	R@1	R@5	R@10	R@1	R@5	R@10
CONCH	0.0052	0.0411	0.1021	0.0108	0.0215	0.0323
GloSAGE	0.1694	0.2723	0.3292	0.0215	0.0645	0.0753
KEEP	0.0109	0.0415	0.0863	0.0108	0.0323	0.0538
PLIP	0.0081	0.1210	0.2126	0.0108	0.0430	0.0753

Appendix

Table S4: Retrieval performance on PAS-stained images

	Image-to-Text			Text-to-Image
Model	R@1	R@5	R@10	R@1	R@5	R@10
CONCH	0.0033	0.0658	0.2242	0.0500	0.0500	0.2000
GloSAGE	0.2867	0.4826	0.6007	0.1000	0.1500	0.3000
KEEP	0.0076	0.0397	0.1094	0.0500	0.1500	0.2000
PLIP	0.2671	0.7193	0.8069	0.1000	0.1500	0.2000

Appendix

Table S5: Retrieval performance on Masson-stained images

	Image-to-Text			Text-to-Image
Model	R@1	R@5	R@10	R@1	R@5	R@10
CONCH	0.0062	0.0355	0.0719	0.0392	0.1176	0.1373
GloSAGE	0.1083	0.2233	0.2655	0.0588	0.0784	0.1569
KEEP	0.0110	0.0652	0.1112	0.0000	0.0392	0.0784
PLIP	0.0585	0.2530	0.4528	0.0196	0.0588	0.0784

Appendix

Table S6: Retrieval performance on PASM-stained images

Model	PAS	Masson	PASM
GloSAGE	0.283	0.201	0.179
PLIP	0.106	0.139	0.116
CONCH	0.176	0.083	0.179
KEEP	0.226	0.120	0.188

Appendix

Table S7: Mean hard-negative description-discrimination accuracy by staining. Each lesion-positive image is matched to one correct and five closely related incorrect descriptions; chance top-1 accuracy is 0.1667. Means assign equal weight to each evaluated lesion (12 PAS, 8 Masson and 11 PASM), including additional stress-test lesions beyond the 27 canonical zero-shot recognition tasks. GloPath is omitted here to match the models displayed in Fig. 2 C.

Lesion	GloSAGE	PLIP	CONCH	KEEP
PAS-MH	0.002	0.008	0.003	0.021
PAS-PAS( $+$ )	0.244	0.000	0.962	0.000
PAS-PAS( $-$ )	0.827	0.013	0.000	0.680
PAS-Crum	0.161	0.000	0.000	0.032
PAS-ADH	0.014	0.000	0.000	0.000
PAS-SS	0.092	0.618	0.066	0.895

Appendix

Table S8: Lesion-wise hard-negative top-1 accuracy for PAS. The final row is the unweighted mean across all 12 evaluated lesion targets; PAS-SFN is an additional stress-test target beyond the standard 11-lesion recognition panel.

Lesion	GloSAGE	PLIP	CONCH	KEEP
MT-MH	0.333	0.004	0.046	0.007
MT-ADH	0.051	0.000	0.000	0.000
MT-SS	0.233	0.000	0.067	0.900
MT-SFN	0.030	0.104	0.060	0.000
MT-Cre	0.184	0.000	0.000	0.000
MT-GS	0.375	1.000	0.492	0.055

Appendix

Table S9: Lesion-wise hard-negative top-1 accuracy for Masson. The final row is the unweighted mean across all eight evaluated lesion targets; MT-Crum and MT-MN are additional stress-test targets beyond the standard six-lesion recognition panel.

Lesion	GloSAGE	PLIP	CONCH	KEEP
PASM-MH	0.030	0.000	0.000	0.017
PASM-PAS( $-$ )	0.384	0.000	0.233	0.178
PASM-ADH	0.000	0.000	0.000	0.000
PASM-SS	0.029	0.000	0.088	0.794
PASM-SFN	0.263	0.289	0.237	0.184
PASM-MN	0.343	0.051	0.000	0.056

Appendix

Table S10: Lesion-wise hard-negative top-1 accuracy for PASM. The final row is the unweighted mean across all 11 evaluated lesion targets; PASM-Crum is an additional stress-test target beyond the standard ten-lesion recognition panel.

Model	MH	PAS( $+$ )	PAS( $-$ )	Crum	ADH	SS	MN	MP	EP	Cre	GS	Overall
GloSAGE	0.523	0.664	0.884	0.551	0.441	0.696	0.721	0.660	0.564	0.662	0.710	0.651
PLIP	0.502	0.482	0.361	0.515	0.501	0.607	0.513	0.537	0.409	0.428	0.577	0.492
CONCH	0.455	0.495	0.529	0.477	0.360	0.343	0.560	0.581	0.645	0.808	0.334	0.509
KEEP	0.589	0.383	0.517	0.583	0.396	0.512	0.554	0.682	0.310	0.604	0.708	0.529

Appendix

Table S11: Intervention-discrimination ROC-AUC under PAS staining. The score is the change in image–text similarity between the original and intervened descriptions; ROC-AUC evaluates its separation of lesion-positive versus lesion-negative images. It is not a difference between two diagnostic ROC-AUC values. Each lesion entry averages its valid interventions; Overall averages valid interventions across lesions.

Model	MH	ADH	SS	SFN	Cre	GS	Overall
GloSAGE	0.518	0.488	0.654	0.529	0.683	0.621	0.584
PLIP	0.551	0.411	0.584	0.345	0.561	0.741	0.507
CONCH	0.404	0.421	0.332	0.604	0.637	0.336	0.481
KEEP	0.454	0.514	0.480	0.570	0.502	0.578	0.475

Appendix

Table S13: Intervention-discrimination ROC-AUC under Masson staining. See Supplementary Table S11 for the score definition and aggregation.

Model	MH	PAS( $-$ )	ADH	SS	SFN	MN	MP	EP	Cre	GS	Overall
GloSAGE	0.527	0.840	0.437	0.665	0.636	0.595	0.545	0.616	0.721	0.549	0.613
PLIP	0.562	0.436	0.432	0.381	0.414	0.505	0.496	0.543	0.548	0.527	0.485
CONCH	0.417	0.653	0.428	0.334	0.506	0.617	0.497	0.732	0.773	0.456	0.532
KEEP	0.451	0.620	0.556	0.480	0.674	0.608	0.580	0.375	0.725	0.559	0.559

Appendix

Table S14: Intervention-discrimination ROC-AUC under PASM staining. See Supplementary Table S11 for the score definition and aggregation.

Model	MH	PAS(+)	PAS( $-$ )	Crum	ADH	SS	MN	MP	EP	Cre	GS
GloSAGE	0.930 $\pm$ 0.005	0.954 $\pm$ 0.002	1.000 $\pm$ 0.000	0.922 $\pm$ 0.024	0.815 $\pm$ 0.014	0.967 $\pm$ 0.004	0.957 $\pm$ 0.001	0.968 $\pm$ 0.003	0.975 $\pm$ 0.002	0.989 $\pm$ 0.003	0.997 $\pm$ 0.001
GloPath	0.909 $\pm$ 0.038	0.955 $\pm$ 0.003	1.000 $\pm$ 0.000	0.887 $\pm$ 0.037	0.779 $\pm$ 0.033	0.962 $\pm$ 0.006	0.949 $\pm$ 0.013	0.965 $\pm$ 0.005	0.968 $\pm$ 0.015	0.980 $\pm$ 0.005	0.997 $\pm$ 0.001
UNI	0.879 $\pm$ 0.092	0.936 $\pm$ 0.008	1.000 $\pm$ 0.000	0.842 $\pm$ 0.108	0.762 $\pm$ 0.016	0.924 $\pm$ 0.019	0.947 $\pm$ 0.006	0.952 $\pm$ 0.008	0.969 $\pm$ 0.004	0.984 $\pm$ 0.005	0.996 $\pm$ 0.001
CONCH	0.884 $\pm$ 0.032	0.934 $\pm$ 0.004	0.997 $\pm$ 0.001	0.777 $\pm$ 0.157	0.720 $\pm$ 0.027	0.919 $\pm$ 0.011	0.923 $\pm$ 0.006	0.954 $\pm$ 0.006	0.965 $\pm$ 0.002	0.963 $\pm$ 0.008	0.993 $\pm$ 0.001
PLIP	0.889 $\pm$ 0.005	0.924 $\pm$ 0.004	0.996 $\pm$ 0.002	0.842 $\pm$ 0.048	0.765 $\pm$ 0.016	0.915 $\pm$ 0.024	0.868 $\pm$ 0.012	0.950 $\pm$ 0.003	0.943 $\pm$ 0.004	0.967 $\pm$ 0.007	0.994 $\pm$ 0.002
KEEP	0.926 $\pm$ 0.002	0.936 $\pm$ 0.005	1.000 $\pm$ 0.000	0.861 $\pm$ 0.020	0.760 $\pm$ 0.015	0.951 $\pm$ 0.012	0.923 $\pm$ 0.011	0.946 $\pm$ 0.004	0.966 $\pm$ 0.011	0.966 $\pm$ 0.005	0.998 $\pm$ 0.000

Appendix

Table S15: Performance of lesion recognition across all the models (PAS)

Model	MH	ADH	SS	SFN	Cre	GS
GloSAGE	0.962 $\pm$ 0.005	0.763 $\pm$ 0.024	0.948 $\pm$ 0.013	0.986 $\pm$ 0.004	0.963 $\pm$ 0.002	0.997 $\pm$ 0.001
GloPath	0.961 $\pm$ 0.003	0.750 $\pm$ 0.050	0.910 $\pm$ 0.054	0.982 $\pm$ 0.007	0.974 $\pm$ 0.018	0.995 $\pm$ 0.003
UNI	0.952 $\pm$ 0.005	0.773 $\pm$ 0.080	0.868 $\pm$ 0.060	0.990 $\pm$ 0.002	0.986 $\pm$ 0.005	0.995 $\pm$ 0.000
CONCH	0.928 $\pm$ 0.007	0.704 $\pm$ 0.038	0.709 $\pm$ 0.076	0.985 $\pm$ 0.001	0.983 $\pm$ 0.008	0.993 $\pm$ 0.001
PLIP	0.929 $\pm$ 0.003	0.708 $\pm$ 0.099	0.710 $\pm$ 0.042	0.984 $\pm$ 0.006	0.960 $\pm$ 0.018	0.992 $\pm$ 0.002
KEEP	0.930 $\pm$ 0.005	0.723 $\pm$ 0.008	0.786 $\pm$ 0.077	0.984 $\pm$ 0.006	0.984 $\pm$ 0.004	0.992 $\pm$ 0.002

Appendix

Table S16: Performance of lesion recognition across all the models (Masson)

Model	MH	PAS( $-$ )	ADH	SS	SFN	MN	MP	EP	Cre	GS
GloSAGE	0.928 $\pm$ 0.001	0.998 $\pm$ 0.001	0.762 $\pm$ 0.030	0.961 $\pm$ 0.010	0.986 $\pm$ 0.002	0.946 $\pm$ 0.012	0.953 $\pm$ 0.001	0.975 $\pm$ 0.001	0.988 $\pm$ 0.000	0.996 $\pm$ 0.002
GloPath	0.928 $\pm$ 0.004	0.997 $\pm$ 0.001	0.768 $\pm$ 0.038	0.945 $\pm$ 0.018	0.982 $\pm$ 0.004	0.944 $\pm$ 0.008	0.955 $\pm$ 0.008	0.972 $\pm$ 0.009	0.982 $\pm$ 0.010	0.995 $\pm$ 0.001
UNI	0.924 $\pm$ 0.003	0.998 $\pm$ 0.001	0.747 $\pm$ 0.022	0.943 $\pm$ 0.031	0.977 $\pm$ 0.011	0.929 $\pm$ 0.013	0.969 $\pm$ 0.001	0.971 $\pm$ 0.006	0.983 $\pm$ 0.009	0.992 $\pm$ 0.001
CONCH	0.912 $\pm$ 0.005	0.993 $\pm$ 0.004	0.695 $\pm$ 0.013	0.933 $\pm$ 0.018	0.945 $\pm$ 0.014	0.894 $\pm$ 0.014	0.943 $\pm$ 0.007	0.953 $\pm$ 0.006	0.938 $\pm$ 0.004	0.990 $\pm$ 0.002
PLIP	0.905 $\pm$ 0.005	0.993 $\pm$ 0.002	0.790 $\pm$ 0.049	0.902 $\pm$ 0.034	0.946 $\pm$ 0.005	0.920 $\pm$ 0.009	0.953 $\pm$ 0.005	0.971 $\pm$ 0.002	0.968 $\pm$ 0.019	0.991 $\pm$ 0.002
KEEP	0.912 $\pm$ 0.006	0.984 $\pm$ 0.021	0.749 $\pm$ 0.057	0.914 $\pm$ 0.012	0.955 $\pm$ 0.015	0.927 $\pm$ 0.002	0.957 $\pm$ 0.004	0.972 $\pm$ 0.007	0.944 $\pm$ 0.041	0.989 $\pm$ 0.001

Appendix

Table S17: Performance of lesion recognition across all the models (PASM)

Lesion	GloPath	UNI	CONCH	PLIP	KEEP
MH	+22.38	+41.96	+39.41	+36.34	+5.18
PAS(+)	-1.70	+28.35	+30.52	+39.78	+27.62
PAS(-)	+100.00	+100.00	+100.00	+100.00	+100.00
Crum	+30.81	+50.71	+64.99	+50.81	+43.83
ADH	+15.98	+22.09	+33.75	+21.06	+22.69
SS	+13.59	+56.85	+59.77	+61.70	+32.94

Appendix

Table S18: Relative reduction in ROC-AUC deficit of GloSAGE compared with baseline models on PAS staining.

Lesion	GloPath	UNI	CONCH	PLIP	KEEP
MH	+2.41	+21.48	+47.76	+46.80	+46.20
ADH	+5.19	-4.37	+20.03	+18.75	+14.28
SS	+41.96	+60.13	+81.95	+81.91	+75.44
SFN	+23.17	-34.63	+7.61	+12.74	+13.30
Cre	-42.09	-160.00	-113.14	+7.66	-123.56
GS	+48.72	+45.71	+59.87	+66.23	+63.45

Appendix

Table S19: Relative reduction in ROC-AUC deficit of GloSAGE compared with baseline models on Masson staining.

Lesion	GloPath	UNI	CONCH	PLIP	KEEP
MH	-0.40	+5.79	+18.59	+24.33	+18.39
PAS(-)	+32.15	+29.44	+75.50	+74.80	+88.92
ADH	-2.35	+6.25	+22.00	-12.90	+5.47
SS	+28.37	+31.15	+41.98	+60.01	+54.35
SFN	+20.90	+37.48	+74.11	+73.49	+68.23
MN	+3.17	+23.46	+48.47	+31.72	+25.54

Appendix

Table S20: Relative reduction in ROC-AUC deficit of GloSAGE compared with baseline models on PASM staining.

Task	GloSAGE	GloPath	UNI	PLIP	CONCH	KEEP
PAS-MH-Grading	0.897 $\pm$ 0.005	0.896 $\pm$ 0.004	0.894 $\pm$ 0.004	0.873 $\pm$ 0.005	0.861 $\pm$ 0.004	0.890 $\pm$ 0.009
MT-MH-Grading	0.920 $\pm$ 0.009	0.907 $\pm$ 0.011	0.901 $\pm$ 0.007	0.883 $\pm$ 0.006	0.852 $\pm$ 0.012	0.856 $\pm$ 0.015
PASM-MH-Grading	0.892 $\pm$ 0.005	0.893 $\pm$ 0.006	0.877 $\pm$ 0.005	0.885 $\pm$ 0.011	0.834 $\pm$ 0.014	0.877 $\pm$ 0.012
PAS-Cre-Grading	0.976 $\pm$ 0.016	0.982 $\pm$ 0.008	0.946 $\pm$ 0.011	0.949 $\pm$ 0.031	0.932 $\pm$ 0.028	0.964 $\pm$ 0.017
MT-Cre-Grading	0.961 $\pm$ 0.015	0.908 $\pm$ 0.046	0.915 $\pm$ 0.032	0.840 $\pm$ 0.028	0.881 $\pm$ 0.074	0.922 $\pm$ 0.043
PASM-Cre-Grading	0.925 $\pm$ 0.016	0.937 $\pm$ 0.016	0.936 $\pm$ 0.029	0.915 $\pm$ 0.005	0.892 $\pm$ 0.036	0.892 $\pm$ 0.054

Appendix

Table S21: Performance comparison of models on lesion grading tasks. Values are ROC-AUC (mean

\pm

standard deviation).

Model	MH	PAS(+)	PAS(-)	Crum	ADH	SS	MN	MP	EP	Cre	GS
GloSAGE	0.9196	0.8462	0.9900	0.7863	0.8127	0.9142	0.8807	0.9919	0.9995	0.9680	0.9974
GloPath	0.8969	0.8565	0.9830	0.7467	0.7904	0.9205	0.9194	0.9817	0.9987	0.9542	0.9966
UNI	0.8869	0.8226	0.9858	0.7458	0.7706	0.8660	0.8914	0.9620	0.9937	0.9462	0.9976
PLIP	0.8591	0.8126	0.9645	0.7649	0.7739	0.8722	0.8791	0.9052	0.9956	0.9363	0.9963
CONCH	0.8373	0.7979	0.9570	0.6999	0.7478	0.8177	0.8686	0.9426	0.9641	0.9074	0.9954
KEEP	0.8788	0.8365	0.9653	0.7769	0.7560	0.8099	0.8685	0.9534	0.9930	0.8851	0.9966

Appendix

Table S22: Performance on 11 PAS-related lesion recognition tasks on GloPro.

Model	ROC-AUC	F1-score	Accuracy
GloSAGE	$\mathbf{0.9873\pm 0.0085}$	$\mathbf{0.9444\pm 0.0274}$	$\mathbf{0.9445\pm 0.0273}$
GloPath	$0.9777\pm 0.0077$	$0.9335\pm 0.0087$	$0.9336\pm 0.0087$
UNI	$0.9582\pm 0.0216$	$0.8925\pm 0.0378$	$0.8933\pm 0.0375$
PLIP	$0.9123\pm 0.0102$	$0.7562\pm 0.0531$	$0.7676\pm 0.0476$
CONCH	$0.8740\pm 0.0270$	$0.7870\pm 0.0175$	$0.7905\pm 0.0171$
KEEP	$0.8908\pm 0.0330$	$0.8056\pm 0.0299$	$0.8081\pm 0.0275$

Appendix

Table S23: Performance comparison on disease classification on RenalXJ (deep feature based). Values are mean

\pm

sample standard deviation across 3 folds. The task involves IgA nephropathy versus membranous nephropathy classification.

Model	ROC-AUC	F1-score	Accuracy
GloSAGE	$\mathbf{0.7358\pm 0.0310}$	$\mathbf{0.7054\pm 0.0078}$	$\mathbf{0.7352\pm 0.0193}$
GloPath	$0.6915\pm 0.0209$	$0.6361\pm 0.0137$	$0.6718\pm 0.0393$
UNI	$0.7013\pm 0.0358$	$0.6571\pm 0.0104$	$0.7206\pm 0.0090$
PLIP	$0.6968\pm 0.0126$	$0.6007\pm 0.0342$	$0.6974\pm 0.0036$
CONCH	$0.6681\pm 0.0104$	$0.6312\pm 0.0539$	$0.7169\pm 0.0228$
KEEP	$0.7080\pm 0.0457$	$0.6022\pm 0.0213$	$0.7048\pm 0.0047$

Appendix

Table S24: Performance comparison on injury grading on RenalXJ (deep feature based). Values are mean

\pm

sample standard deviation across 3 folds. The task involves three kidney injury grades (0, 1, and 2).

Model	F1-score	AUC	Acc
GloSAGE	0.8657 $\pm$ 0.0017	0.9334 $\pm$ 0.0258	0.8667 $\pm$ 0.0015
GloPath	0.8432 $\pm$ 0.0585	0.9458 $\pm$ 0.0088	0.8448 $\pm$ 0.0559
UNI	0.8367 $\pm$ 0.0117	0.9291 $\pm$ 0.0193	0.8377 $\pm$ 0.0118
CONCH	0.7839 $\pm$ 0.0038	0.9004 $\pm$ 0.0185	0.7852 $\pm$ 0.0045
PLIP	0.7703 $\pm$ 0.0353	0.8966 $\pm$ 0.0171	0.7743 $\pm$ 0.0316
KEEP	0.7504 $\pm$ 0.0699	0.8797 $\pm$ 0.0221	0.7632 $\pm$ 0.0549

Appendix

Table S25: Performance comparison on disease classification on RenalXJ (lesion-based).

Model	AUC	Acc
GloSAGE	0.7433 $\pm$ 0.0406	0.7259 $\pm$ 0.0231
GloPath	0.7690 $\pm$ 0.0609	0.6963 $\pm$ 0.0039
UNI	0.7252 $\pm$ 0.0518	0.6974 $\pm$ 0.0044
CONCH	0.7533 $\pm$ 0.0785	0.7398 $\pm$ 0.0058
PLIP	0.7322 $\pm$ 0.0674	0.6963 $\pm$ 0.0039
KEEP	0.7238 $\pm$ 0.0546	0.6926 $\pm$ 0.0073

Appendix

Table S26: Performance comparison on injury classification on RenalXJ (lesion-based).

Model	MH	PAS( $+$ )	PAS( $-$ )	Crum	ADH	SS	MN	MP	EP	Cre	GS
GloSAGE	0.9296 $\pm$ 0.0048	0.9539 $\pm$ 0.0019	1.0000 $\pm$ 0.0000	0.9220 $\pm$ 0.0239	0.8146 $\pm$ 0.0139	0.9673 $\pm$ 0.0038	0.9574 $\pm$ 0.0014	0.9685 $\pm$ 0.0031	0.9746 $\pm$ 0.0018	0.9893 $\pm$ 0.0026	0.9967 $\pm$ 0.0006
w/o EDGE	0.9222 $\pm$ 0.0042	0.9503 $\pm$ 0.0016	1.0000 $\pm$ 0.0000	0.9034 $\pm$ 0.0254	0.7965 $\pm$ 0.0057	0.9555 $\pm$ 0.0189	0.9476 $\pm$ 0.0077	0.9674 $\pm$ 0.0019	0.9749 $\pm$ 0.0012	0.9823 $\pm$ 0.0024	0.9943 $\pm$ 0.0018
w/o KEYS	0.9226 $\pm$ 0.0049	0.9479 $\pm$ 0.0042	1.0000 $\pm$ 0.0000	0.8813 $\pm$ 0.0418	0.7955 $\pm$ 0.0138	0.9667 $\pm$ 0.0128	0.9554 $\pm$ 0.0023	0.9666 $\pm$ 0.0034	0.9721 $\pm$ 0.0026	0.9818 $\pm$ 0.0086	0.9964 $\pm$ 0.0005
w/o EDGE and KEYS	0.9192 $\pm$ 0.0097	0.9416 $\pm$ 0.0080	1.0000 $\pm$ 0.0000	0.8779 $\pm$ 0.0336	0.7789 $\pm$ 0.0090	0.9558 $\pm$ 0.0217	0.9521 $\pm$ 0.0053	0.9675 $\pm$ 0.0050	0.9688 $\pm$ 0.0070	0.9817 $\pm$ 0.0085	0.9965 $\pm$ 0.0003
GloSAGE-P	0.9260 $\pm$ 0.0082	0.9484 $\pm$ 0.0031	0.9998 $\pm$ 0.0003	0.8685 $\pm$ 0.0823	0.7692 $\pm$ 0.0096	0.9297 $\pm$ 0.0332	0.9464 $\pm$ 0.0085	0.9676 $\pm$ 0.0015	0.9512 $\pm$ 0.0229	0.9814 $\pm$ 0.0072	0.9946 $\pm$ 0.0027
Vanilla visual pretraining (GloPath)	0.909 $\pm$ 0.038	0.955 $\pm$ 0.003	1.000 $\pm$ 0.000	0.887 $\pm$ 0.037	0.779 $\pm$ 0.033	0.962 $\pm$ 0.009	0.949 $\pm$ 0.013	0.965 $\pm$ 0.005	0.968 $\pm$ 0.015	0.980 $\pm$ 0.005	0.997 $\pm$ 0.001

Appendix

Table S27: Ablation results for lesion recognition on PAS-stained images. Values are reported as mean

\pm

standard deviation. Bold indicates the highest mean among the five GloSAGE configurations, including ties; the GloPath baseline is excluded from this comparison.

Model	MH	ADH	SS	SFN	Cre	GS
GloSAGE	0.9622 $\pm$ 0.0045	0.7630 $\pm$ 0.0243	0.9475 $\pm$ 0.0130	0.9859 $\pm$ 0.0035	0.9633 $\pm$ 0.0016	0.9972 $\pm$ 0.0005
w/o EDGE	0.9643 $\pm$ 0.0027	0.7174 $\pm$ 0.0711	0.9212 $\pm$ 0.0277	0.9907 $\pm$ 0.0076	0.9535 $\pm$ 0.0268	0.9941 $\pm$ 0.0025
w/o KEYS	0.9597 $\pm$ 0.0080	0.7177 $\pm$ 0.0341	0.9409 $\pm$ 0.0272	0.9927 $\pm$ 0.0036	0.9428 $\pm$ 0.0517	0.9959 $\pm$ 0.0012
w/o EDGE and KEYS	0.9648 $\pm$ 0.0052	0.7710 $\pm$ 0.0928	0.9077 $\pm$ 0.0518	0.9877 $\pm$ 0.0164	0.9491 $\pm$ 0.0602	0.9955 $\pm$ 0.0002
GloSAGE-P	0.9509 $\pm$ 0.0034	0.7324 $\pm$ 0.0883	0.7966 $\pm$ 0.0270	0.9772 $\pm$ 0.0101	0.9345 $\pm$ 0.0351	0.9947 $\pm$ 0.0019
Vanilla visual pretraining (GloPath)	0.961 $\pm$ 0.003	0.750 $\pm$ 0.050	0.910 $\pm$ 0.054	0.982 $\pm$ 0.007	0.974 $\pm$ 0.018	0.995 $\pm$ 0.003

Appendix

Table S28: Ablation results for lesion recognition on Masson-stained images. Values are reported as mean

\pm

standard deviation. Bold indicates the highest mean among the five GloSAGE configurations; the GloPath baseline is excluded from this comparison.

Model	MH	PAS( $-$ )	ADH	SS	SFN	MN	MP	EP	Cre	GS
GloSAGE	0.9281 $\pm$ 0.0012	0.9982 $\pm$ 0.0007	0.7623 $\pm$ 0.0299	0.9609 $\pm$ 0.0102	0.9857 $\pm$ 0.0017	0.9456 $\pm$ 0.0119	0.9535 $\pm$ 0.0014	0.9747 $\pm$ 0.0014	0.9880 $\pm$ 0.0004	0.9961 $\pm$ 0.0020
w/o EDGE	0.9251 $\pm$ 0.0030	0.9974 $\pm$ 0.0007	0.7455 $\pm$ 0.0430	0.9344 $\pm$ 0.0167	0.9789 $\pm$ 0.0032	0.9278 $\pm$ 0.0095	0.9472 $\pm$ 0.0032	0.9635 $\pm$ 0.0210	0.9882 $\pm$ 0.0044	0.9935 $\pm$ 0.0019
w/o KEYS	0.9207 $\pm$ 0.0101	0.9964 $\pm$ 0.0020	0.7358 $\pm$ 0.0129	0.9138 $\pm$ 0.0536	0.9838 $\pm$ 0.0045	0.9408 $\pm$ 0.0168	0.9543 $\pm$ 0.0093	0.9588 $\pm$ 0.0093	0.9884 $\pm$ 0.0007	0.9919 $\pm$ 0.0006
w/o EDGE and KEYS	0.9286 $\pm$ 0.0053	0.9974 $\pm$ 0.0013	0.7830 $\pm$ 0.0522	0.9486 $\pm$ 0.0059	0.9818 $\pm$ 0.0078	0.9438 $\pm$ 0.0033	0.9470 $\pm$ 0.0111	0.9701 $\pm$ 0.0045	0.9887 $\pm$ 0.0046	0.9932 $\pm$ 0.0017
GloSAGE-P	0.9198 $\pm$ 0.0063	0.9969 $\pm$ 0.0022	0.7466 $\pm$ 0.1000	0.8750 $\pm$ 0.0677	0.9714 $\pm$ 0.0047	0.9253 $\pm$ 0.0063	0.9439 $\pm$ 0.0060	0.9572 $\pm$ 0.0092	0.9790 $\pm$ 0.0041	0.9907 $\pm$ 0.0026
Vanilla visual pretraining (GloPath)	0.928 $\pm$ 0.004	0.997 $\pm$ 0.001	0.768 $\pm$ 0.038	0.945 $\pm$ 0.018	0.982 $\pm$ 0.004	0.944 $\pm$ 0.008	0.955 $\pm$ 0.008	0.972 $\pm$ 0.009	0.982 $\pm$ 0.010	0.995 $\pm$ 0.001

Appendix

Table S29: Ablation results for lesion recognition on PASM-stained images. Values are reported as mean

\pm

standard deviation. Bold indicates the highest mean among the five GloSAGE configurations; the GloPath baseline is excluded from this comparison.

Model	ROUGE-1	ROUGE-2	ROUGE-L
GloSAGE	0.7906	0.7143	0.7791
PLIP	0.7744	0.6950	0.7660
CONCH	0.7859	0.7099	0.7776
KEEP	0.7757	0.6963	0.7671

Appendix

Table S30: Performance of image-conditioned glomerular description generation. ROUGE-1, ROUGE-2, and ROUGE-L are recall-based overlap scores against knowledge-graph-derived reference descriptions, averaged across PAS, Masson, and PASM.

Pathology Foundation Models

Pathology foundation models are pretrained AI systems that learn reusable tissue-image representations for diagnosis, biomarker discovery, and clinical workflows.

See all 91 papers

GloPath: An Entity-Centric Foundation Model for Glomerular Lesion Assessment and Clinicopathological Insights
Mar 3, 2026Qiming He, Jing Li, Tian Guan +26Pathology Foundation Models Histopathology Image Classification
Institute of Biopharmaceutical and Health Engineering, Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · Interdisciplinary Institute for Medical Engineering, Fuzhou University, Fuzhou, China · Medical Optical Technology R&D Center, Research Institute of Tsinghua, Pearl River Delta, Guangzhou, China +6
From Classification to Cross-Modal Understanding: Leveraging Vision-Language Models for Fine-Grained Renal Pathology
Nov 15, 2025Zhenhao Guo, Rachit Saluja, Tianyu Shi +13Medical VLMs VLM Adaptation
New York University, New York, NY 10012, USA · Cornell Tech, New York, NY 10044, USA · Vanderbilt University, Nashville, TN, 37235, USA +4
ALICE: Learning a General-Purpose Pathology Foundation Model from Vision, Vision-Language, and Slide-Level Experts
Jul 10, 2026Jiawen Li, Tian Guan, Huijuan Shi +5Computational Pathology Unified Multimodal Models
Institute of Biopharmaceutical and Health Engineering, Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · Department of Pathology, The First Affiliated Hospital of Sun Yat-sen University, Guangzhou, China · Department of Engineering Science, University of Oxford, Oxford, UK +1

Mar 3, 2026J/K move · Enter open · S save

GloPath: An Entity-Centric Foundation Model for Glomerular Lesion Assessment and Clinicopathological Insights

Qiming He, Jing Li, Tian Guan, Yifei Ma+25

Institute of Biopharmaceutical and Health Engineering, Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · Interdisciplinary Institute for Medical Engineering, Fuzhou University, Fuzhou, China · Medical Optical Technology R&D Center, Research Institute of Tsinghua, Pearl River Delta, Guangzhou, China+6

Pathology Foundation Models Histopathology Image Classification

Open paper AskPDF arXiv

Glomerular pathology is central to the diagnosis and prognosis of renal diseases, yet the heterogeneity of glomerular morphology and fine-grained lesion patterns remain challenging for current AI approaches. We present GloPath, an entity-centric foundation model trained on over one million glomeruli extracted from 14,049 renal biopsy specimens using multi-scale and multi-view self-supervised learning. GloPath addresses two major challenges in nephropathology: glomerular lesion assessment and clinicopathological insights discovery. For lesion assessment, GloPath was benchmarked across three independent cohorts on 52 tasks, including lesion recognition, grading, few-shot classification, and cross-modality diagnosis-outperforming state-of-the-art methods in 42 tasks (80.8%). In the large-scale real-world study, it achieved an ROC-AUC of 91.51% for lesion recognition, demonstrating strong robustness in routine clinical settings. For clinicopathological insights, GloPath systematically revealed statistically significant associations between glomerular morphological parameters and clinical indicators across 224 morphology-clinical variable pairs, demonstrating its capacity to connect tissue-level pathology with patient-level outcomes. Together, these results position GloPath as a scalable and interpretable platform for glomerular lesion assessment and clinicopathological discovery, representing a step toward clinically translatable AI in renal pathology.

Abstract

Pathology Foundation Models

Explore similar work

GloPath: An Entity-Centric Foundation Model for Glomerular Lesion Assessment and Clinicopathological Insights

From Classification to Cross-Modal Understanding: Leveraging Vision-Language Models for Fine-Grained Renal Pathology

ALICE: Learning a General-Purpose Pathology Foundation Model from Vision, Vision-Language, and Slide-Level Experts