Core Concepts
RetMIL introduces a retentive mechanism to improve WSI analysis performance.
Abstract
The article introduces RetMIL, a method for histopathological whole slide image (WSI) classification. It addresses challenges faced by Transformer-based MIL methods, such as high memory consumption and slow inference speed. RetMIL processes WSI sequences hierarchically, updating tokens through retention mechanisms at local and global levels. Experiments on CAMELYON, BRACS, and LUNG datasets show that RetMIL achieves state-of-the-art performance with reduced computational overhead. The proposed method enhances model interpretability and outperforms Transformer-based models across different sequence lengths.
Stats
RetMIL surpasses TransMIL by 3.18% in F1-score on the CAMELYON dataset.
In the BRACS dataset, RetMIL leads by 1.52% compared to CLAM-MB.
RetMIL outperforms Transformer-based models at different sequence lengths.
Quotes
"Our proposed RetMIL achieves lower memory cost and higher throughput while exhibiting competitive performance."
"RetMIL significantly improves model throughput compared to Transformer-based methods."
"Our observation reveals that RetMIL can better widen the gap between distinct categories while minimizing the separation among patches belonging to the same category."