Feature Optimization-Based Multiple Instance Learning for Whole Slide Image Classification

Manual examination of whole slide images (WSIs) remains the “gold standard” for cancer diagnosis. However, traditional pathology workflows relying on visual inspection by pathologists suffer from inefficiency and poor reproducibility, failing to meet growing clinical demands. While deep learning-based whole slide image classification techniques have gained prominence for their high accuracy and efficiency, critical limitations persist. To address WSI noise interference and domain adaptation challenges, we propose LaKMIL for whole slide image classification. This model enhances classification performance by refining feature representation quality through learnable kernel fusion and attention-guided instance selection. Evaluations on two public datasets (Camelyon16 and TCGA-Lung) demonstrate that LaKMIL achieves state-of-the-art accuracy and F1 scores using only image-level labels, outperforming baseline MIL methods by $3.7-5.2 \%$ in cross-domain testing scenarios.