Hierarchical attention-driven multiple instance learning for clear cell renal cell carcinoma grading and staging in digital pathology

doi:10.12122/j.issn.1673-4254.2026.05.22

Abstract

Abstract:

Objective To develop a hierarchical attention-driven multiple instance learning (HA-MIL) framework for precise grading and staging of clear cell renal cell carcinoma (ccRCC). Methods Whole-slide images (WSIs) and the corresponding clinicopathological labels were obtained from 504 patients from the TCGA-KIRC database. The proposed HA-MIL framework processes WSIs by segmenting them into patches, leverages a hierarchical attention mechanism to discern the significance of instances across different levels and incorporates a dynamic gated fusion module to refine feature aggregation. This approach mitigates the dependency on subjective evaluations by pathologists. Results Experimental results showed that HA-MIL achieved an accuracy of (87.35±0.72)%, an AUC of 0.95, and an F1-score of (86.41±0.63)% in the staging task. For the grading task, the model attained an accuracy of (85.82±0.60)%, an AUC of 0.93, and an F1-score of (86.00±0.65)%. HA-MIL demonstrated significantly higher accuracy and robustness compared to the baseline methods including SVM, DSMIL, and ABMIL. Ablation studies confirmed the contribution of the hierarchical attention mechanism, which improved the accuracy and F1-score by 3.2% and 2.2%, respectively, outperforming the conventional attention mechanisms. Conclusion The HA-MIL framework exhibits superior performance for grading and staging of ccRCC, thus offering a novel, intelligent method for pathological image analysis of ccRCC in the clinical setting.

Key words: clear cell renal cell carcinoma, deep learning, multiple instance learning, attention mechanism, grading and staging

Jianing XU, Yixiao MAO, Yu ZHANG. Hierarchical attention-driven multiple instance learning for clear cell renal cell carcinoma grading and staging in digital pathology[J]. Journal of Southern Medical University, 2026, 46(5): 1175-1184.

Figures/Tables 8

Fig.1 Framework of the proposed method. A: Thresholding patching. B: Patch feature extraction. C: Feature augmentation. D: Global sparse attention. E: Local attention.

Fig.2 Patch segmentation and feature extraction process.

Fig.3 Performance evaluation experiment. A: Diagnostic performance in grade 1, 2, 3, and 4. B: Diagnostic performance in stage 1, 2, 3, and 4. SVM: Support vector machine; DSMIL: Dual-stream multiple instance learning; ABMIL: Attention-based deep multiple instance learning; HA-MIL: Hierarchical attention-driven multiple instance learning.

Tab.1 Comparative performance evaluation of the grading and staging algorithms (Mean±SD)

Method	Accuracy (%)	F1-score (%)	FLOPs (G)	Parameters (M)
a:Diagnostic performance in Grade (Grade-1, 2, 3, 4)
SVM	84.77±1.29	85.12±1.21	0.86±0.12	0.53
DSMIL	83.03±2.17	83.75±2.04	5.34±0.28	23.87
ABMIL	83.78±1.40	84.22±1.33	5.92±0.37	25.34
HA-MIL (ours)	85.82±0.60**	86.00±0.65	6.85±0.49	28.63
b:Diagnostic performance in stage (Stage-1, 2, 3, 4)
SVM	86.03±0.88	86.21±0.82	-	-
DSMIL	82.45±1.21	81.25±1.35	-	-
ABMIL	84.16±0.68	84.60±0.65	-	-
HA-MIL (ours)	87.35±0.72*	86.41±0.63	-	-

Fig.4 ROC curves and confusion matrix. A: Grading and staging results with the HA-MIL model. B: Comparison before and after feature enhancement.

Tab.2 Ablation experiment of the feature enhancement module (Mean±SD)

Augmentation strategy	Accuracy (%)	F1-score (%)
HA-MIL (none)	50.51±1.25	31.13±1.40
HA-MIL (scaling)	67.42±1.88	70.15±2.11
HA-MIL (rotation)	64.78±2.05	67.89±2.33
HA-MIL (noise)	57.91±1.97	60.12±2.40
HA-MIL (traditional)	72.37±1.75	80.71±2.05
HA-MIL (proposed)	71.94±2.25	79.83±2.40

Tab.3 Ablation experiment of the gating module (Mean±SD)

Model	Accuracy (%)	F1-score (%)
HA-MIL (L)	83.46±1.17	84.22±1.33
HA-MIL (G)	76.55±8.55	66.55±16.95
HA-MIL (frozen g)	83.91±2.83	82.77±3.51
HA-MIL (proposed)	87.35±0.72	86.41±0.63

Fig.5 Visualization of attention heatmaps and high, medium, low attention score patches for each phase (HE staining, original magnification: ×20). The heatmap colors range from blue to red, indicating feature values from low to high probability.

References 31

[1]	Hsieh JJ, Purdue MP, Signoretti S, et al. Renal cell carcinoma[J]. Nat Rev Dis Primers, 2017, 3: 17009. doi：10.1016/j.celrep.2017.02.074
[2]	Powles T, Albiges L, Bex A, et al. Renal cell carcinoma: ESMO Clinical Practice Guideline for diagnosis, treatment and follow-up[J]. Ann Oncol, 2024, 35(8): 692-706. doi：10.1016/j.annonc.2024.05.537
[3]	Ljungberg B, Albiges L, Abu-Ghanem Y, et al. European association of urology guidelines on renal cell carcinoma: the 2019 update[J]. Eur Urol, 2019, 75(5): 799-810.
[4]	Chen RJ, Lu MY, Wang JW, et al. Pathomic fusion: an integrated framework for fusing histopathology and genomic features for cancer diagnosis and prognosis[J]. IEEE Trans Med Imaging, 2022, 41(4): 757-70. doi：10.1109/tmi.2020.3021387
[5]	Litjens G, Kooi T, Bejnordi BE, et al. A survey on deep learning in medical image analysis[J]. Med Image Anal, 2017, 42: 60-88. doi：10.1016/j.media.2017.07.005
[6]	Xiao Q, Yi XP, Guan X, et al. Validation of the World Health Organization/International Society of Urological Pathology grading for Chinese patients with clear cell renal cell carcinoma[J]. Transl Androl Urol, 2020, 9(6): 2665-74. doi：10.21037/tau-20-799
[7]	Campanella G, Hanna MG, Geneslaw L, et al. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images[J]. Nat Med, 2019, 25(8): 1301-9. doi：10.1038/s41591-019-0508-1
[8]	Cai JX, Liu MT, Zhang Q, et al. Renal cancer detection: fusing deep and texture features from histopathology images[J]. Biomed Res Int, 2022, 2022: 9821773. doi：10.1155/2022/9821773
[9]	Zhou K, Wei LH. Grading prediction of kidney renal clear cell carcinoma by deep learning[C]//2022 2nd International Conference on Bioinformatics and Intelligent Computing. Harbin China. ACM, 2022: 37-42. doi：10.1145/3523286.3524510
[10]	Yao J, Wei L, Hao PP, et al. Application of artificial intelligence model in pathological staging and prognosis of clear cell renal cell carcinoma[J]. Discov Oncol, 2024, 15(1): 545. doi：10.1007/s12672-024-01437-8
[11]	Bex A, Abu Ghanem Y, Albiges L, et al. European association of urology guidelines on renal cell carcinoma: the 2025 update[J]. Eur Urol, 2025, 87(6): 683-96.
[12]	Lu MY, Williamson DFK, Chen TY, et al. Data-efficient and weakly supervised computational pathology on whole-slide images[J]. Nat Biomed Eng, 2021, 5(6): 555-70. doi：10.1038/s41551-020-00682-w
[13]	Verma V, Lamb A, Beckham C, et al. Manifold mixup: better representations by interpolating hidden states[EB/OL]. arXiv, 2018. . doi：10.48550/arXiv.1806.05236
[14]	Zhang L, Wang XS, Yang D, et al. Generalizing deep learning for medical image segmentation to unseen domains via deep stacked transformation[J]. IEEE Trans Med Imaging, 2020, 39(7): 2531-40. doi：10.1109/TMI.2020.2973595
[15]	Wu YC, Ge ZY, Zhang DH, et al. Mutual consistency learning for semi-supervised medical image segmentation[J]. Med Image Anal, 2022, 81: 102530. doi：10.1016/j.media.2022.102530
[16]	Chen RJ, Lu MY, Williamson DFK, et al. Pan-cancer integrative histology-genomic analysis via multimodal deep learning[J]. Cancer Cell, 2022, 40(8): 865-78.e6. doi：10.1016/j.ccell.2022.07.004
[17]	Ilse M, Tomczak J, Welling M. Attention-based deep multiple instance learning[J/OL]. arXiv, 2018.
[18]	Shao ZC, Bian H, Chen Y, et al. TransMIL: transformer based correlated multiple instance learning for whole slide image classification[J/OL]. arXiv, 2021. .
[19]	Im YH, Park SH, Lee SC. HDA-net: H&E and RGB dual attention network for nuclei instance segmentation[J]. IEEE Access, 2024, 12: 56622-32. doi：10.1109/access.2024.3390726
[20]	Deng RN, Cui C, Remedios LW, et al. Cross-scale multi-instance learning for pathological image diagnosis[J]. Med Image Anal, 2024, 94: 103124. doi：10.1016/j.media.2024.103124
[21]	Xie X, Luo YB, Peng HW, et al. RTop-K: ultra-fast row-wise top-K selection for neural network acceleration on GPUs[C]//International Conference on Learning Representations, 2024.
[22]	Lin TY, Goyal P, Girshick R, et al. Focal loss for dense object detection[J]. IEEE Trans Pattern Anal Mach Intell, 2020, 42(2): 318-27. doi：10.1109/tpami.2018.2858826
[23]	Cortes C, Vapnik V. Support-vector networks[J]. Mach Learn, 1995, 20(3): 273-97. doi：10.1007/bf00994018
[24]	Li B, Li Y, Eliceiri KW. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning[J]. Conf Comput Vis Pattern Recognit Workshops, 2021, 2021: 14318-28. doi：10.1109/CVPR46437.2021.01409
[25]	Mobadersany P, Yousefi S, Amgad M, et al. Predicting cancer outcomes from histology and genomics using convolutional networks[J]. Proc Natl Acad Sci U S A, 2018, 115(13): E2970-9. doi：10.1073/pnas.1717139115
[26]	Lu MY, Chen TY, Williamson DFK, et al. AI-based pathology predicts origins for cancers of unknown primary[J]. Nature, 2021, 594(7861): 106-10. doi：10.1038/s41586-021-03512-4
[27]	Ehteshami Bejnordi B, Veta M, Johannes van Diest P, et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer[J]. JAMA, 2017, 318(22): 2199-210. doi：10.1001/jama.2017.14580
[28]	Ghassemi M, Oakden-Rayner L, Beam AL. The false hope of current approaches to explainable artificial intelligence in health care[J]. Lancet Digit Health, 2021, 3(11): e745-50. doi：10.1016/s2589-7500(21)00208-9
[29]	van der Laak J, Litjens G, Ciompi F. Deep learning in histopathology: the path to the clinic[J]. Nat Med, 2021, 27(5): 775-84. doi：10.1038/s41591-021-01343-4
[30]	Saillard C, Schmauch B, Laifa O, et al. Predicting survival after hepatocellular carcinoma resection using deep learning on histological slides[J]. Hepatology, 2020, 72(6): 2000-13. doi：10.1002/hep.31207
[31]	Bulten W, Pinckaers H, van Boven H, et al. Automated deep-learning system for Gleason grading of prostate cancer using biopsies: a diagnostic study[J]. Lancet Oncol, 2020, 21(2): 233-41. doi：10.1016/S1470-2045(19)30739-9