Enhancing Whole Slide Image Classification in Renal Cell Carcinoma via Swin Transformer-Based Multiple Instance...

Area of Science:

• Digital pathology
• Computational oncology
• Artificial intelligence in medicine

Background:

• Renal cell carcinoma (RCC) has distinct histologic subtypes impacting prognosis and treatment.
• Accurate subtyping of RCC is crucial for effective patient management.
• Current classification methods can be complex and time-consuming.

Purpose of the Study:

• To evaluate a weakly supervised, slide-level AI model for classifying RCC subtypes.
• To assess the performance of Swin-CLAM, a modified CLAM model using Swin-Tiny Transformer, for RCC subtype classification.
• To investigate the utility of advanced AI in improving the accuracy of renal cancer subtyping.

Main Methods:

• Utilized 928 whole-slide images (WSIs) from TCGA-RCC for clear cell (ccRCC), papillary (pRCC), and chromophobe (chRCC) subtypes.
• Employed Swin-CLAM, integrating a Swin-Tiny Transformer patch encoder with the CLAM-SB aggregation module.
• WSIs were processed into 256x256 patches and classified using slide-level labels in a five-fold cross-validation.

Main Results:

• Swin-CLAM achieved a macro-averaged AUC of 0.976, accuracy of 94.8%, and macro-F1 of 0.940 on the TCGA-RCC dataset.
• The model demonstrated the largest performance improvement for chromophobe RCC classification.
• Qualitative analyses included attention heatmaps and t-SNE plots for exploratory insights.

Conclusions:

• Stronger patch-level representations, like those from Swin-Tiny Transformers, enhance CLAM-based RCC subtype classification.
• The proposed Swin-CLAM model shows significant potential for automated RCC subtyping.
• Further validation, calibration, and domain-shift analysis are necessary for clinical implementation.