TSMIL: Transformer-based structured low-rank end-to-end multi-instance learning network for renal cell carcinoma classification in whole-slide images

The pathological classification of renal cell carcinoma (RCC) is a critical indicator of its accurate diagnosis, treatment, and prognosis. Pathologists typically focus on a single subtype when determining classifications, whereas existing multiple instance learning approaches often lack the global modeling of instance-level features and fail to capture contextual dependencies. When multiple instances with large semantic disparities are projected into the same latent space, structural information loss or the dilution of critical pathological features may occur, thereby limiting classification performance. To address these challenges, we propose an end-to-end Transformer-based structured low-rank multiple instance learning framework, termed TSMIL, for RCC classification. Specifically, we introduce the multilayer spatial feature module to enhance morphological feature representation and the structured low-rank block to embed high-dimensional features into a low-rank structure, effectively capturing contextual information and exploring the latent semantic potential of pathological features in a sparse representation space. Extensive experiments demonstrate that our proposed TSMIL achieves a mean accuracy of 92.98% and an AUC of 0.9818, outperforming other state-of-the-art methods. Overall, our framework exhibits superior practicality and robustness in RCC pathology grading tasks.