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

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

cs.CVMar 3, 2026

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

Authors: Qiming He, Jing Li, Tian Guan, Yifei Ma, Zimo Zhao, Yanxia Wang, Hongjing Chen, Yingming Xu, +21 more

Organizations: 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 · Department of Pathology, School of Basic Medicine and Xijing Hospital, Fourth Military Medical University, Xi’an, China · Department of Engineering Science, University of Oxford, Oxford, UK · Department of Pathology, Ankang Central Hospital, Ankang, China · Peng Cheng Laboratory, Shenzhen, China · Jinan Inspur Data Technology Co., Ltd., Jinan, China · Department of Pathology, Xijing Hospital, Fourth Military Medical University, Xi’an, China

Abstract

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.

Figures & tables

Figure 1: Overview of GloPath. a, Dataset curation. The pathological images are obtained from seven cohorts including XJ-Light-1, XJ-Light-2, XJ-IF, AIDPATH-G, XJ-GIO and KPMP-G, where WSI is processed by entity detection. b, Entity-centric self-supervised pretraining based on multi-scale multi-view. c, Lesion assessment. d, Clinicopathological correlation analysis.

Figure 2: Overview of the performance of lesion assessment. a, The best model on each lesion assessment task across lesion recognition, lesion grading, cross-modality diagnosis, and few-shot-based diagnosis. Refer to the legend in (b). b, Comparison of the distribution of the results of lesion recognition.

Figure 3: Comparison of results of glomerular lesion analysis. a, Comparison of the distribution of F1 score of lesion recognition. b-d, Comparison of results on PAS, MT and PASM stainings. The primary vertical axis corresponds to the F1 score of each model. The secondary vertical axis corresponds to Min-Gain and Max-Gain, which are, respectively, the minimum and maximum performance gains of GloPath over all other models. Below the chart are the visualization results, where rows 1-3 are original images and the results of GloPath and UNI, respectively.

Figure 4: Comparison of results of cross-modality diagnosis. a, Comparison of preformance on cross-modality diagnosis. b, Visualization of attention map of GloPath. Row 1-2 indicate region classification and row 3-4 indicate pattern classification. Column 1 indicates the original IF images and column 2-4 indicate the attention map from three attention heads. c-d, Validation loss curves for region and pattenr respectively.

Figure 5: Comparison of results of few-shot based classification. a, Performance of few-shot based classification on AIDPATH-G. b, Distribution of embedding on AIDPATH-G using t-SNE. c, Visualization of Channel-wise embedding on AIDPATH-G. d, Results of few-shot for IF image classification. Row 1-2 indicate results for region, pattern respectively. Column 1-4 indicate results for different methods including LR, MLP, RF, and PTL respectively.

Figure 6: Overview of clinicopathological correlation analysis. a, Visualization of bow and tuft segmentation and morphological parameter definition. b-c, Comparison of distribution of significance levels on XJ-Light-1 and KPMP-G.

Figure 7: Details of clinicopathological correlation analysis. a-f, The p-values of correlation analysis. a-b , c-d and e-f refer to GloPath, UNI and RenalPath, respectively. The left and right column refer to XJ-Light-1 and KPMP-G respectively. g-h show examples from XJ-Light-1 and KPMP-G.

Appendix figures & tables63 assets

Supplementary material from the paper’s appendix.

Appendix

Figure S1: Comparison of IoU of all the models on morphological segmentation, which shows the results of bow segmentaion on XJ-Light and tuft segmentation on XJ-GIO.

Figure S2: Comparison of segmentation visulization of all the models. a, Performance of bow segmentaion. Column 1-10 indicate original iamegs, ground truth, results of CONCH, GloPath, RenalPath, ImageNetPre, RandomInit, PLIP, UNI and nnU-Net. The masks represent the target areas. b, Performance of tuft segmentaion.

Figure S3: Performance of all the compared models on few-shot leanring on AIDPATH-G with confidence intervals.

Figure S4: Performance of all the compared models on few-shot leanring on XJ-IF with confidence intervals. Row 1-2 indicate region and pattern classification tasks respectively.

Figure S5: Clinicopathological correlation analysis on XJ-Light-1. The x-axis represents clinical parameters and their respective groups, while the y-axis shows the statistical distribution of pathological phenotypes within each group. Exact P values are provided in Table S16-S22.

Figure S6: Clinicopathological correlation analysis on KPMP-G. The definitions of the x-axis and y-axis are the same as in Fig. S5 . Exact P values are provided in Table S25-S33.

Figure S7: Examples of the XJ-Light dataset. a, Examples of slides from 12 levels. b, Examples of glomeruli from four stainings.

Figure S8: Interface of the Kidney Label Tool which is used for lesion annotation.

Figure S9: Examples of the XJ-IF dataset with different deposition regions and patterns and with diverse markers.

Figure S10: Examples of the AIDPATH-G dataset. a, Renal pathology images and their annotations. b, Examples of normal and sclerotic glomeruli.

Figure S11: Examples of slides and extracted glomeruli with different stainings in the KPMP-G dataset.

Figure S12: Error analysis of GloPath on XJ-CLI. For each lesion type, we present one false negative and one false positive example, with GloPath’s predicted probability for the lesion displayed below each glomerular image (a probability $p>0.5$ indicates the presence of the lesion).

Figure S13: UMAP projection across different weights and imaging domains.

Figure S14: Interpretability study of GloPath. a, Comparison between heatmaps generated from GloPath predictions and annotations by nephrologists (GS). Columns 1-4 indicate original images, original images with annotations, heatmaps, and overlays. b, Comparison between heatmaps generated from GloPath predictions and annotations by nephrologists (Cre). c-d, Comparison of mean attention values between target (lesion) and non-target (non-lesion) regions for GS and Cre, respectively. KS-test results (KS D and p-value) are shown below the violin plots.

Lesion Recognition
PAS-MH recognition	Recognition of thylakoid hyperplasia in PAS staining
PAS-PAS(+) recognition	Recognition PAS positivity in PAS staining
PAS-PAS(-) recognition	Recognition PAS negativity in PAS staining
PAS-Crum recognition	Recognition of the presence of ischemic crumpling in PAS staining
PAS-ADH recognition	Recognition the presence of adhesions in PAS staining
PAS-SS recognition	Recognition the presence of segmental sclerosis in PAS staining

Appendix

Table S1: Details of the 52 lesion assessment tasks.

Lesion	MH	PAS(+)	PAS(-)	Crum	ADH	SS	MN	MP	EP	Cre	GS	None of above lesions
n	1987	1441	24	119	511	90	611	5	1	132	954	9330

Appendix

Table S2: Number of glomeruli with each type of lesion in XJ-CLI.

Method	MH	PAS(+)	PAS(-)	Crum	ADH	SS	MN	MP	EP	Cre	GS
RandomInit	0.720	0.629	0.374	0.803	0.275	0.324	0.834	0.664	0.321	0.720	0.882
ImagenetPre	0.653	0.675	0.643	0.699	0.343	0.323	0.722	0.727	0.154	0.588	0.945
CONCH	0.782	0.870	0.990	0.683	0.686	0.942	0.896	0.923	0.950	0.932	0.955
PLIP	0.789	0.862	0.954	0.649	0.619	0.794	0.898	0.827	0.839	0.935	0.966
UNI	0.832	0.877	0.999	0.924	0.591	0.919	0.948	0.919	0.927	0.962	0.978
RenalPath	0.820	0.873	0.994	0.844	0.791	0.852	0.891	0.919	0.920	0.928	0.979

Appendix

Table S3: Comparison of methods on lesion recognition in PAS staining (F1 score). The bold value indicates the best model.

Method	MH	ADH	SS	SFN	Cre	GS
RandomInit	0.735	0.441	0.515	0.768	0.636	0.790
ImagenetPre	0.663	0.623	0.393	0.725	0.508	0.791
CONCH	0.897	0.948	0.951	0.978	0.895	0.974
PLIP	0.891	0.966	0.907	0.853	0.885	0.955
UNI	0.883	0.895	0.973	0.974	0.955	0.975
RenalPath	0.917	0.928	0.895	0.974	0.911	0.977

Appendix

Table S4: Comparison of methods on lesion recognition in MT staining (F1 score). The bold value indicates the best model.

Method	MH	PAS(-)	ADH	SS	SFN	MN	MP	EP	Cre	GS
RandomInit	0.530	0.548	0.624	0.695	0.738	0.791	0.805	0.890	0.798	0.890
ImagenetPre	0.610	0.381	0.530	0.863	0.753	0.408	0.698	0.848	0.907	0.838
CONCH	0.864	0.890	0.954	0.947	0.949	0.871	0.922	0.966	0.946	0.950
PLIP	0.815	0.981	0.890	0.896	0.814	0.893	0.923	0.946	0.963	0.950
UNI	0.877	0.986	0.894	0.964	0.966	0.855	0.920	0.969	0.958	0.959
RenalPath	0.887	0.982	0.774	0.919	0.960	0.862	0.901	0.957	0.946	0.966

Appendix

Table S5: Comparison of methods on lesion recognition in PASM staining (F1 score). The bold value indicates the best model.

Method	RandomInit	ImageNetPre	CONCH	PLIP	UNI	RenalPath	GloPath
ROC-AUC	0.878	0.893	0.992	0.990	0.999	0.995	0.999

Appendix

Table S6: Comparison of methods on AIDPATH-G under full supervision. The bold value indicates the best model.

Method	k=1	k=5	k=10	k=25	k=100
RandomInit	0.578	0.667	0.691	0.724	0.777
ImageNetPre	0.573	0.665	0.690	0.727	0.783
CONCH	0.606	0.771	0.828	0.876	0.960
PLIP	0.674	0.836	0.886	0.923	0.966
UNI	0.719	0.877	0.912	0.959	0.987
RenalPath	0.619	0.803	0.841	0.899	0.968

Appendix

Table S7: Comparison of methods on AIDPATH-G using SVM-based few-shot learning. The bold value indicates the best model.

Method	k=1	k=5	k=10	k=25	k=100
RandomInit	0.578	0.677	0.781	0.835	0.886
ImageNetPre	0.573	0.682	0.783	0.839	0.890
CONCH	0.607	0.801	0.857	0.923	0.978
PLIP	0.674	0.853	0.910	0.946	0.979
UNI	0.716	0.896	0.934	0.969	0.989
RenalPath	0.619	0.823	0.884	0.933	0.976

Appendix

Table S8: Comparison of methods on AIDPATH-G using LR-based few-shot learning. The bold value indicates the best model.

Method	k=1	k=5	k=10	k=25	k=100
RandomInit	0.577	0.662	0.762	0.821	0.887
ImageNetPre	0.573	0.675	0.768	0.831	0.893
CONCH	0.575	0.785	0.845	0.917	0.973
PLIP	0.662	0.861	0.908	0.941	0.976
UNI	0.717	0.873	0.915	0.964	0.984
RenalPath	0.641	0.813	0.891	0.933	0.975

Appendix

Table S9: Comparison of methods on AIDPATH-G using MLP-based few-shot learning. The bold value indicates the best model.

Method	k=1	k=5	k=10	k=25	k=100
RandomInit	0.567	0.681	0.742	0.798	0.845
ImageNetPre	0.544	0.681	0.751	0.801	0.849
CONCH	0.515	0.762	0.828	0.875	0.931
PLIP	0.474	0.837	0.892	0.922	0.953
UNI	0.542	0.872	0.909	0.947	0.979
RenalPath	0.514	0.785	0.857	0.904	0.950

Appendix

Table S10: Comparison of methods on AIDPATH-G using RF-based few-shot learning. The bold value indicates the best model.

Method	k=1	k=5	k=10	k=25	k=100
RandomInit	0.578	0.661	0.696	0.712	0.710
ImageNetPre	0.573	0.661	0.692	0.708	0.707
CONCH	0.606	0.783	0.827	0.838	0.856
PLIP	0.674	0.854	0.886	0.903	0.910
UNI	0.718	0.884	0.901	0.919	0.933
RenalPath	0.619	0.796	0.823	0.840	0.835

Appendix

Table S11: Comparison of methods on AIDPATH-G using PTL-based few-shot learning. The bold value indicates the best model.

Method	PAS-MH-Grading	MT-MH-Grading	PASM-MH-Grading	PAS-Cre-Grading	MT-Cre-Grading	PASM-Cre-Grading	PAS-MP-Grading	PASM-MP-Grading	PAS-EP-Grading	PASM-EP-Grading
RandomInit	0.584	0.704	0.716	0.824	0.827	0.796	0.854	0.815	0.766	0.849
ImageNetPre	0.584	0.548	0.736	0.857	0.856	0.809	0.854	0.864	0.851	0.900
CONCH	0.700	0.866	0.849	0.927	0.954	0.931	0.910	0.849	0.932	0.969
PLIP	0.673	0.864	0.804	0.911	0.822	0.926	0.894	0.870	0.909	0.902
UNI	0.736	0.884	0.833	0.943	0.945	0.913	0.920	0.915	0.954	0.961
RenalPath	0.729	0.886	0.846	0.933	0.943	0.916	0.896	0.906	0.933	0.924

Appendix

Table S12: Comparison of methods on lesion grading (ROC-AUC). The bold value indicates the best model.

Method	Bow	Tuft
RandomInit	0.754	0.736
ImageNetPre	0.754	0.738
CONCH	0.754	0.768
PLIP	0.754	0.826
UNI	0.835	0.887
RenalPath	0.824	0.876

Appendix

Table S13: Comparison of methods on semantic segmentation of the glomerular bow and tuft. The bold value indicates the best model.

Significance Level		**
GloPath	18	12	7
UNI	21	8	8
RenalPath	14	9	6

Appendix

Table S14: Comparison of significance level of pathomics-data mining of XJ-Light-1 for GloPath, UNI, and RenalPath.

Significance Level		**
GloPath	24	12	7
UNI	16	10	10
RenalPath	16	13	6

Appendix

Table S15: Comparison of significance level of pathomics-data mining of KPMP-G for GloPath, UNI, and RenalPath.

Clinic Parameter: Gender
Pathology Phenotype	GloPath	UNI	RenalPath
AreaBow_mean	0.0066 (0.1738)	0.0125 (0.1636)	0.0125 (0.1636)
AreaBow_med	0.0184 (0.1572)	0.0365 (0.1463)	0.0342 (0.1452)
AreaTuft_mean	0.0027 (0.1870)	0.0006 (0.1847)	0.0032 (0.2065)
AreaTuft_med	0.0456 (0.1410)	0.0371 (0.1661)	0.0107 (0.1449)
Ratio_mean	0.1220 (0.1210)	0.1731 (0.0812)	0.5433 (0.1131)

Appendix

Table S16: Significance and effect size of clinicopathological correlation analysis for Gender on XJ-Light-1 (p-value (D))

Clinic Parameter: Age
Pathology Phenotype	GloPath	UNI	RenalPath
AreaBow_mean	0.6849 (0.0791)	0.5131 (0.0661)	0.7777 (0.0827)
AreaBow_med	0.9145 (0.0556)	0.9608 (0.0669)	0.7636 (0.0500)
AreaTuft_mean	0.7726 (0.0664)	0.9135 (0.0669)	0.7631 (0.0556)
AreaTuft_med	0.2771 (0.1003)	0.5695 (0.0889)	0.4179 (0.0787)
Ratio_mean	0.0127 (0.1621)	0.1341 (0.1625)	0.0123 (0.1180)

Appendix

Table S17: Significance and effect size of clinicopathological correlation analysis for Age on XJ-Light-1 (p-value (D))

Clinic Parameter: Creatinine
Pathology Phenotype	GloPath	UNI	RenalPath
AreaBow_mean	0.1154 (0.1472)	0.3181 (0.1462)	0.1200 (0.1175)
AreaBow_med	0.0513 (0.1672)	0.0661 (0.1984)	0.0117 (0.1612)
AreaTuft_mean	0.1389 (0.1422)	0.1460 (0.1586)	0.0734 (0.1409)
AreaTuft_med	0.0548 (0.1656)	0.0524 (0.1969)	0.0127 (0.1667)
Ratio_mean	0.0419 (0.1718)	0.2449 (0.1469)	0.1169 (0.1259)

Appendix

Table S18: Significance and effect size of clinicopathological correlation analysis for Creatinine on XJ-Light-1 (p-value (D))

Clinic Parameter: IgA (Binary)
Pathology Phenotype	GloPath	UNI	RenalPath
AreaBow_mean	0.0007 (0.2240)	0.0002 (0.2156)	0.0012 (0.2438)
AreaBow_med	0.0091 (0.1838)	0.0098 (0.2059)	0.0023 (0.1824)
AreaTuft_mean	0.0041 (0.1969)	0.0027 (0.2062)	0.0023 (0.2033)
AreaTuft_med	0.0829 (0.1407)	0.0333 (0.1484)	0.0583 (0.1598)
Ratio_mean	0.5024 (0.0913)	0.3824 (0.0978)	0.4158 (0.1006)

Appendix

Table S19: Significance and effect size of clinicopathological correlation analysis for IgA (Binary) on XJ-Light-1 (p-value (D))

Clinic Parameter: IgA Lee Score
Pathology Phenotype	GloPath	UNI	RenalPath
AreaBow_mean	0.3151 (0.0008)	0.3200 (0.0033)	0.1986 (0.0008)
AreaBow_med	0.0028 (0.0266)	0.0034 (0.0320)	0.0010 (0.0254)
AreaTuft_mean	0.1219 (0.0060)	0.1274 (0.0083)	0.0805 (0.0058)
AreaTuft_med	0.0001 (0.0463)	0.0001 (0.0546)	0.0000 (0.0446)
Ratio_mean	0.0122 (0.0185)	0.0131 (0.0133)	0.0316 (0.0181)

Appendix

Table S20: Significance and effect size of clinicopathological correlation analysis for IgA Lee Score on XJ-Light-1 (p-value (

\epsilon^{2}

))

Clinic Parameter: Disease
Pathology Phenotype	GloPath	UNI	RenalPath
AreaBow_mean	0.0000 (0.1013)	0.0000 (0.1051)	0.0000 (0.1038)
AreaBow_med	0.0000 (0.0652)	0.0000 (0.0688)	0.0000 (0.0172)
AreaTuft_mean	0.0000 (0.0972)	0.0000 (0.1099)	0.0000 (0.0980)
AreaTuft_med	0.0003 (0.0476)	0.0002 (0.0552)	0.0001 (0.0501)
Ratio_mean	0.0080 (0.0267)	0.0258 (0.0317)	0.0036 (0.0193)

Appendix

Table S21: Significance and effect size of clinicopathological correlation analysis for Disease on XJ-Light-1 (p-value (

\epsilon^{2}

))

Clinic Parameter: Lesion
Pathology Phenotype	GloPath	UNI	RenalPath
AreaBow_mean	0.0963 (0.0073)	0.0819 (0.0062)	0.1184 (0.0082)
AreaBow_med	0.8887 (0.0000)	0.8090 (0.0000)	0.9395 (0.0000)
AreaTuft_mean	0.4036 (0.0000)	0.3238 (0.0002)	0.3569 (0.0007)
AreaTuft_med	0.8736 (0.0000)	0.9567 (0.0000)	0.8917 (0.0000)
Ratio_mean	0.0273 (0.0141)	0.0359 (0.0152)	0.0226 (0.0126)

Appendix

Table S22: Significance and effect size of clinicopathological correlation analysis for Lesion on XJ-Light-1 (p-value (

\epsilon^{2}

))

Pathology Phenotype	$R^{2}$	coef (G)	p (G)	coef (A)	coef (D)	p (D)	coef (P)	p (P)
Proposed_AreaBow_mean	0.1071	23.1147	0.0135	1.5866	-45.8914	0.0001	-0.0004	0.0476
Proposed_AreaBow_med	0.1188	23.4835	0.0113	1.5234	-44.6479	0.0001	-0.0005	0.0031
Proposed_AreaTuft_mean	0.1202	25.1494	0.0070	1.5210	-42.4377	0.0002	-0.0008	0.0023
Proposed_AreaTuft_med	0.1204	23.5983	0.0108	1.4747	-44.5297	0.0001	-0.0006	0.0022
Proposed_Ratio_mean	0.1167	22.6231	0.0145	1.4406	-46.7163	0.0000	-201.0923	0.0050
Proposed_Ratio_med	0.1262	22.6289	0.0139	1.4404	-44.5344	0.0001	-219.8215	0.0006

Appendix

Table S23: Multivariate Regression Analysis of Pathology Phenotypes Associated with Creatinine. Other dependent variables: Gender, Age, Disease Type (IgAN, LN, MN, DN, and MCD). G, A, D, and P indicate gender, age, disease type, and pathology phenotype respectively.

Pathology Phenotype	$R^{2}$	coef (G)	p (G)	coef (A)	p (A)	coef (D)	p (D)	coef (P)	p (P)
Proposed_AreaBow_mean	0.1193	0.1964	0.5911	-0.0126	0.3174	-0.9831	0.1347	0.0000	0.0001
Proposed_AreaBow_med	0.1202	0.1415	0.6986	-0.0149	0.2406	-1.1577	0.0750	0.0000	0.0001
Proposed_AreaTuft_mean	0.1291	0.2466	0.5036	-0.0152	0.2349	-0.9330	0.1560	0.0000	0.0000
Proposed_AreaTuft_med	0.1125	0.1601	0.6600	-0.0158	0.2156	-1.2221	0.0587	0.0000	0.0001
Proposed_Ratio_mean	0.0660	0.1283	0.7201	-0.0134	0.2865	-1.3799	0.0299	-5.5755	0.0170
Proposed_Ratio_med	0.0763	0.1272	0.7241	-0.0133	0.2876	-1.3252	0.0375	-5.6022	0.0042

Appendix

Table S24: Multivariate Regression Analysis of Pathology Phenotypes Associated with Disease Type. Other dependent variables: Gender, Age, Disease Type (IgAN, LN, MN, DN, and MCD).

Clinic Parameter: Enrollment Category
Pathology Phenotype	GloPath	UNI	RenalPath
AreaBow_mean	0.0041 (0.0000)	0.0024 (0.0000)	0.0059 (0.0000)
AreaBow_med	0.6383 (0.0000)	0.5592 (0.0000)	0.6403 (0.0000)
AreaTuft_mean	0.1044 (0.0000)	0.2850 (0.0000)	0.0335 (0.0000)
AreaTuft_med	0.3037 (0.0000)	0.7116 (0.0000)	0.8076 (0.0000)
Ratio_mean	0.5496 (0.0000)	0.0072 (0.0000)	0.0457 (0.0000)

Appendix

Table S25: Significance and effect size of clinicopathological correlation analysis for Enrollment Category on KPMP-G (p-value (

\epsilon^{2}

))

Clinic Parameter: Gender
Pathology Phenotype	GloPath	UNI	RenalPath
AreaBow_mean	0.5660 (0.0869)	0.9777 (0.0884)	0.5440 (0.0515)
AreaBow_med	0.9727 (0.0526)	0.7460 (0.0484)	0.9880 (0.0746)
AreaTuft_mean	0.9923 (0.0465)	0.9671 (0.0563)	0.9516 (0.0537)
AreaTuft_med	0.7014 (0.0777)	0.0800 (0.0909)	0.5088 (0.0142)
Ratio_mean	0.1904 (0.1207)	0.4768 (0.1238)	0.1689 (0.0932)

Appendix

Table S26: Significance and effect size of clinicopathological correlation analysis for Gender on KPMP-G (p-value (D))

Clinic Parameter: Age
Pathology Phenotype	GloPath	UNI	RenalPath
AreaBow_mean	0.0591 (0.0185)	0.0607 (0.0165)	0.0747 (0.0182)
AreaBow_med	0.0125 (0.0313)	0.0189 (0.0268)	0.0217 (0.0280)
AreaTuft_mean	0.0256 (0.0255)	0.0320 (0.0274)	0.0204 (0.0236)
AreaTuft_med	0.0351 (0.0229)	0.1036 (0.0201)	0.0488 (0.0136)
Ratio_mean	0.3937 (0.0006)	0.0583 (0.0000)	0.4656 (0.0186)

Appendix

Table S27: Significance and effect size of clinicopathological correlation analysis for Age on KPMP-G (p-value (

\epsilon^{2}

))

Clinic Parameter: Proteinuria
Pathology Phenotype	GloPath	UNI	RenalPath
AreaBow_mean	0.0127 (0.0223)	0.0189 (0.0157)	0.0319 (0.0195)
AreaBow_med	0.2012 (0.0021)	0.1729 (0.0003)	0.2519 (0.0032)
AreaTuft_mean	0.0140 (0.0216)	0.0367 (0.0081)	0.0904 (0.0147)
AreaTuft_med	0.2247 (0.0012)	0.2766 (0.0000)	0.4925 (0.0000)
Ratio_mean	0.8597 (0.0000)	0.4209 (0.0000)	0.4607 (0.0000)

Appendix

Table S28: Significance and effect size of clinicopathological correlation analysis for Proteinuria on KPMP-G (p-value (

\epsilon^{2}

))

Clinic Parameter: A1c
Pathology Phenotype	GloPath	UNI	RenalPath
AreaBow_mean	0.0001 (0.0574)	0.0003 (0.0506)	0.0003 (0.0523)
AreaBow_med	0.0011 (0.0429)	0.0008 (0.0405)	0.0015 (0.0450)
AreaTuft_mean	0.0007 (0.0453)	0.0007 (0.0440)	0.0009 (0.0458)
AreaTuft_med	0.0051 (0.0320)	0.0132 (0.0344)	0.0035 (0.0252)
Ratio_mean	0.5624 (0.0000)	0.0639 (0.0000)	0.6657 (0.0139)

Appendix

Table S29: Significance and effect size of clinicopathological correlation analysis for A1c on KPMP-G (p-value (

\epsilon^{2}

))

Clinic Parameter: Albuminuria
Pathology Phenotype	GloPath	UNI	RenalPath
AreaBow_mean	0.0002 (0.2975)	0.0004 (0.2859)	0.0005 (0.2885)
AreaBow_med	0.0467 (0.1907)	0.0263 (0.1842)	0.0596 (0.2049)
AreaTuft_mean	0.0003 (0.2949)	0.0010 (0.2476)	0.0037 (0.2717)
AreaTuft_med	0.1949 (0.1494)	0.1971 (0.1343)	0.3000 (0.1490)
Ratio_mean	0.8728 (0.0804)	0.8504 (0.1389)	0.2643 (0.0828)

Appendix

Table S30: Significance and effect size of clinicopathological correlation analysis for Albuminuria on KPMP-G (p-value (D))

Clinic Parameter: Diabetes History
Pathology Phenotype	GloPath	UNI	RenalPath
AreaBow_mean	0.0004 (0.2446)	0.0006 (0.2297)	0.0012 (0.2394)
AreaBow_med	0.0085 (0.1964)	0.0158 (0.1466)	0.0939 (0.1851)
AreaTuft_mean	0.0239 (0.1767)	0.1065 (0.1485)	0.0869 (0.1435)
AreaTuft_med	0.1842 (0.1292)	0.3436 (0.1231)	0.2290 (0.1103)
Ratio_mean	0.5747 (0.0914)	0.9438 (0.1464)	0.0946 (0.0609)

Appendix

Table S31: Significance and effect size of clinicopathological correlation analysis for Diabetes History on KPMP-G (p-value (D))

Clinic Parameter: Hypertension History
Pathology Phenotype	GloPath	UNI	RenalPath
AreaBow_mean	0.3550 (0.1368)	0.4629 (0.1098)	0.6268 (0.1252)
AreaBow_med	0.3560 (0.1367)	0.4251 (0.1368)	0.3550 (0.1290)
AreaTuft_mean	0.9289 (0.0768)	0.9085 (0.1069)	0.6563 (0.0817)
AreaTuft_med	0.1961 (0.1596)	0.1222 (0.1559)	0.2163 (0.1752)
Ratio_mean	0.0075 (0.2485)	0.0041 (0.3143)	0.0003 (0.2609)

Appendix

Table S32: Significance and effect size of clinicopathological correlation analysis for Hypertension History on KPMP-G (p-value (D))

Clinic Parameter: eGFR
Pathology Phenotype	GloPath	UNI	RenalPath
AreaBow_mean	0.9834 (0.0000)	0.9294 (0.0000)	0.8818 (0.0000)
AreaBow_med	0.0972 (0.0086)	0.1308 (0.0100)	0.0784 (0.0067)
AreaTuft_mean	0.2920 (0.0015)	0.0550 (0.0042)	0.1933 (0.0123)
AreaTuft_med	0.0240 (0.0177)	0.0041 (0.0239)	0.0093 (0.0291)
Ratio_mean	0.0006 (0.0412)	0.0000 (0.0642)	0.0000 (0.0844)

Appendix

Table S33: Significance and effect size of clinicopathological correlation analysis for eGFR on KPMP-G (p-value (

\epsilon^{2}

))

Method	RandomInit	ImageNetPre	CONCH	PLIP	UNI	RenalPath	GloPath
Region	0.548	0.508	0.850	0.836	0.917	0.914	0.942
Pattern	0.473	0.479	0.801	0.849	0.928	0.912	0.944

Appendix

Table S34: Comparison of methods on XJ-IF for cross-modality diagnosis based on full supervision.

Method	k=1	k=5	k=10	k=25	k=100
RandomInit	-	-	-	0.501	-
ImageNetPre	-	-	-	-	-
CONCH	-	-	-	-	0.823
PLIP	-	-	-	0.561	0.848
UNI	-	-	0.704	0.859	0.929
RenalPath	-	-	0.657	0.685	0.862

Appendix

Table S35: Comparison of methods on deposition region classification on XJ-IF using LR-based few-shot learning. The bold value indicates the best model and the hyphen mean the metrics value is lower than 0.5 and not applicable.

Method	k=1	k=5	k=10	k=25	k=100
RandomInit	-	0.500	0.503	-	0.505
ImageNetPre	-	-	0.503	0.500	0.505
CONCH	-	-	-	-	0.804
PLIP	-	-	-	0.520	0.825
UNI	-	-	0.679	0.854	0.922
RenalPath	-	-	0.674	0.694	0.859

Appendix

Table S36: Comparison of methods on deposition pattern classification on XJ-IF using LR-based few-shot learning. The bold value indicates the best model and the hyphen mean the metrics value is lower than 0.5 and not applicable.

Method	k=1	k=5	k=10	k=25	k=100
RandomInit	0.511	0.514	0.521	0.525	0.550
ImageNetPre	0.512	0.512	0.521	0.519	0.544
CONCH	0.560	0.635	0.719	0.794	0.868
PLIP	0.610	0.733	0.797	0.840	0.880
UNI	0.625	0.717	0.830	0.885	0.939
RenalPath	0.660	0.739	0.792	0.852	0.901

Appendix

Table S37: Comparison of methods on deposition region classification on XJ-IF using MLP-based few-shot learning. The bold value indicates the best model and the hyphen mean the metrics value is lower than 0.5 and not applicable.

Method	k=1	k=5	k=10	k=25	k=100
RandomInit	0.500	0.507	0.510	0.516	0.538
ImageNetPre	0.500	0.507	0.507	0.512	0.532
CONCH	0.536	0.663	0.695	0.753	0.852
PLIP	0.663	0.758	0.762	0.779	0.864
UNI	0.618	0.787	0.822	0.868	0.924
RenalPath	0.662	0.761	0.790	0.817	0.893

Appendix

Table S38: Comparison of methods on deposition pattern classification on XJ-IF using MLP-based few-shot learning. The bold value indicates the best model and the hyphen mean the metrics value is lower than 0.5 and not applicable.

Method	k=1	k=5	k=10	k=25	k=100
RandomInit	0.511	0.516	0.525	0.532	0.553
ImageNetPre	0.512	0.515	0.518	0.513	0.507
CONCH	0.556	0.624	0.705	0.788	0.869
PLIP	0.608	0.730	0.796	0.830	0.876
UNI	0.626	0.706	0.815	0.879	0.939
RenalPath	0.664	0.749	0.786	0.845	0.898

Appendix

Table S39: Comparison of methods on deposition region classification on XJ-IF using RF-based few-shot learning. The bold value indicates the best model and the hyphen mean the metrics value is lower than 0.5 and not applicable.

Method	k=1	k=5	k=10	k=25	k=100
RandomInit	-	0.509	0.513	0.512	0.514
ImageNetPre	-	0.501	0.508	0.507	0.510
CONCH	0.537	0.659	0.687	0.741	0.852
PLIP	0.658	0.759	0.752	0.774	0.860
UNI	0.612	0.783	0.813	0.864	0.927
RenalPath	0.667	0.769	0.785	0.812	0.889

Appendix

Table S40: Comparison of methods on deposition pattern classification on XJ-IF using RF-based few-shot learning. The bold value indicates the best model and the hyphen mean the metrics value is lower than 0.5 and not applicable.

Method	k=1	k=5	k=10	k=25	k=100
RandomInit	0.506	0.514	0.529	0.541	0.556
ImageNetPre	0.499	0.515	0.531	0.536	0.558
CONCH	0.552	0.610	0.687	0.751	0.841
PLIP	0.592	0.722	0.797	0.827	0.860
UNI	0.593	0.682	0.791	0.850	0.908
RenalPath	0.592	0.721	0.791	0.833	0.863

Appendix

Table S41: Comparison of methods on deposition region classification on XJ-IF using PTL-based few-shot learning. The bold value indicates the best model and the hyphen mean the metrics value is lower than 0.5 and not applicable.

Method	k=1	k=5	k=10	k=25	k=100
RandomInit	0.508	0.508	0.517	0.526	0.559
ImageNetPre	0.503	0.508	0.512	0.520	0.555
CONCH	0.537	0.640	0.660	0.708	0.818
PLIP	0.653	0.757	0.764	0.782	0.842
UNI	0.634	0.759	0.794	0.841	0.907
RenalPath	0.620	0.754	0.771	0.805	0.862

Appendix

Table S42: Comparison of methods on deposition pattern classification on XJ-IF using PTL-based few-shot learning. The bold value indicates the best model and the hyphen mean the metrics value is lower than 0.5 and not applicable.

Lesion	MH	PAS(+)	PAS(-)	Crum	ADH	SS	MN	MP	EP	Cre	GS
ROC-AUC	0.897	0.857	0.983	0.747	0.790	0.920	0.919	0.982	0.999	0.954	0.997

Appendix

Table S43: Performance of GloPath in large-scale real-world study.

Staining	MH	PAS(+)	PAS(-)	Crum	ADH	SS	SFN	MN	MP	EP	Cre	GS
PAS	905	479	75	31	222	76	-	101	189	69	110	263
MT	282	-	-	-	39	30	67	-	-	-	38	128
PASM	298	-	73	-	44	34	38	216	132	64	114	164

Appendix

Table S44: Details of the glomerular lesion annotation on XJ-Light-1.

Marker	IgG	IgG2	IgA	IgG4	PLA2R	IgG1	Lambda	C4	C3	C1q	IgG3	Kappa	IgM	AA	Fibrin	IgG-PLA2R1	Albumin	THSD7A	HBsAg	HBeAg	IgG14	7A
Unlabeled	328	4	304	234	136	159	71	9	57	43	4	36	11		3		17	12	23	18	3	4
Labeled	366	31	353	337	178	149	51	6	81	58	22	26	39	10	3	1

Appendix

Table S45: Details of the IF markers in XJ-IF.

Class	Region		Pattern
Class	Capillary	Membrane	Scatted	Diffused
Number	1053	658	747	962

Appendix

Table S46: Details of the annotated images on XJ-IF.

Clinical Variable	Groups
Gender	Male, Female
Age(years old)	>43, $\leq$ 43
Creatinine(umol/L)	>59, $\leq$ 59
IgA(Binary)	non-IgA, IgA
IgA Lee Score	II, III, IV
Disease	IgAN, LN, MN, DN, MCD

Appendix

Table S47: Details of the clinical variables on XJ-Light-1.

Clinical Variable	Groups
Enrollment Category	CKD, AKI, DM-R
Gender	Male, Female
Age(years old)	<30, 30-39, 40-49, 50-59, 60-69, $\geq$ 70
Proteinuria(mg)	<150, 150-500, 500-1000, $\geq$ 1000
A1c(%)	<6.5, 6.5-7.5, 7.5-8.5, $\geq$ 8.5
Albuminuria(mg/L)	<300, $\geq$ 300

Appendix

Table S48: Details of the clinical variables on KPMP-G.

Morphological Variable	Formulation (Glomerular-level)	Definition (case-level)
AreaBow_mean	$S_{bow}$	Mean of all the $S_{bow}$
AreaBow_med	$S_{bow}$	Median of all the $S_{bow}$
AreaTuft_mean	$S_{tuft}$	Mean of all the $S_{tuft}$
AreaTuft_med	$S_{tuft}$	Median of all the $S_{tuft}$
Ratio_mean	$R=\frac{S_{tuft}}{S_{bow}}$	Mean of all the $R$
Ratio_med	$R=\frac{S_{tuft}}{S_{bow}}$	Median of all the $R$

Appendix

Table S49: Details of the morphological variables.

S

P

a

, and

b

refer to the area enclosed by the contour, the perimeter, the major axis, and the minor axis of the ellipse fitted to the contour of the bow or tuft segmented by the model, respectively.

Explore similar work

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
Shortcut Learning in Glomerular AI: Adversarial Penalties Hurt, Entropy Helps
Apr 9, 2026Mohammad Daouk, Jan Ulrich Becker, Neeraja Kambham +3Shortcut Learning Histopathology Image Classification
University of Houston, Houston, TX, USA · University Hospital Cologne, Cologne, Germany · Stanford University, Stanford, CA, USA +1
Seeing Abnormal from Normal: Glomerular Abnormality in Representations of Normal Renal Morphology
Sep 16, 2026Greta Hasko, Rachit Saluja, Tianyu Shi +11Medical Image Anomaly Detection Unsupervised Anomaly Detection
Cornell University, Ithaca, NY 14853, USA · Cornell Tech, New York, NY 10044, USA · Sichuan University, Chengdu, 610207, CN +6

Nov 15, 2025J/K move · Enter open · S save

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

Zhenhao Guo, Rachit Saluja, Tianyu Shi, Tianyuan Yao+12

New York University, New York, NY 10012, USA · Cornell Tech, New York, NY 10044, USA · Vanderbilt University, Nashville, TN, 37235, USA+4

Medical VLMs VLM Adaptation Histopathology Image Classification Fine-Grained Image Classification VLM Calibration

Open paper AskPDF arXiv

Fine-grained glomerular subtyping is essential for kidney biopsy interpretation, yet it presents a harder setting than the broad organ- or disease-level recognition tasks commonly used in vision-language model (VLM) evaluation. In this task, all categories arise within the same renal microanatomic unit, and clinically meaningful subtypes often differ by subtle, overlapping, and continuum-like morphological cues. Clinically valuable labels are also scarce, making fully supervised image-only paradigms difficult to translate to practice. Although VLMs are promising for data-efficient pathology learning, it remains unclear how their cross-modal representations should be adapted and evaluated for such within-structure discrimination under realistic few-shot constraints. To our knowledge, this study establishes the first fine-grained renal pathology benchmark for few-shot VLM adaptation and systematically examines how supervision level, model design, and adaptation strategy shape both classification performance and multimodal representation structure. Beyond accuracy, AUC, and F1, we assess image-text alignment, positive-negative similarity separation, embedding geometry, and calibration (ECE), asking when performance gains reflect more reliable cross-modal understanding. Across renal pathology cohorts, additional supervision consistently improves discrimination, but reliable adaptation is not explained by alignment alone; stronger performance is associated with larger class separation, more coherent embedding organization, and better awareness of confidence-related failure modes under domain variation. Within this retrospective two-cohort evaluation, these findings move assessment beyond coarse classification and accuracy alone and provide comparative evidence to inform model selection and future translational validation. Code: https://github.com/ddrrnn123/glo-vlms

Abstract

Explore similar work

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

Shortcut Learning in Glomerular AI: Adversarial Penalties Hurt, Entropy Helps

Seeing Abnormal from Normal: Glomerular Abnormality in Representations of Normal Renal Morphology