SeLa-MIL: Developing an instance-level classifier via weakly-supervised self-training for whole slide image classification

Pathology image classification is crucial in clinical cancer diagnosis and computer-aided diagnosis. Whole Slide Image (WSI) classification is often framed as a multiple instance learning (MIL) problem due to the high cost of detailed patch-level annotations. Existing MIL methods primarily focus on bag-level classification, often overlooking critical instance-level information, which results in suboptimal outcomes. This paper proposes a novel semi-supervised learning approach, SeLa-MIL, which leverages both labeled and unlabeled instances to improve instance and bag classification, particularly in hard positive instances near the decision boundary.

SeLa-MIL reformulates the traditional MIL problem as a novel semi-supervised instance classification task to effectively utilize both labeled and unlabeled instances. To address the challenge where all labeled instances are negative, we introduce a weakly supervised self-training framework by solving a constrained optimization problem. This method employs global and local constraints on pseudo-labels derived from positive WSI information, enhancing the learning of hard positive instances and ensuring the quality of pseudo-labels. The approach can be integrated into end-to-end training pipelines to maximize the use of available instance-level information.

Comprehensive experiments on synthetic datasets, MIL benchmarks, and popular WSI datasets demonstrate that SeLa-MIL consistently outperforms existing methods in both instance and bag-level classification, with substantial improvements in recognizing hard positive instances. Visualization further highlights the method's effectiveness in pathology regions relevant to cancer diagnosis.

SeLa-MIL effectively addresses key challenges in MIL-based WSI classification by reformulating it as a semi-supervised problem, leveraging both weakly supervised learning and pseudo-labeling techniques. This approach improves classification accuracy and generalization across diverse datasets, making it valuable for pathology image analysis.