Exploring Complexity-Calibrated morphological distribution for whole slide image classification and difficulty-grading

Abstract

Multiple Instance Learning (MIL) is essential for accurate pathological image classification under limited annotations. Global-local morphological modeling approaches have shown promise in whole slide image (WSI) analysis by aligning patches with spatial positions. However, these methods fail to differentiate samples by complexity during morphological distribution construction, treating all samples equally important for model training. This oversight disregards the impact of difficult-to-recognize samples, leading to a morphological fitting bottleneck that hinders the clinical application of deep learning across centers, subtypes, and imaging standards. To address this, we propose Complexity-Calibrated MIL (CoCaMIL) for WSI classification and difficulty grading. CoCaMIL emphasizes the synergistic effects between morphological distribution and key complexity factors, including blur, tumor size, coloring style, brightness, and stain. Specifically, we developed an image–text contrastive pretraining framework to jointly learn multiple complexity factors, enhancing morphological distribution fitting. Additionally, to reduce the tendency to focus on difficult samples overly, we introduce a complexity calibration method, which forms a distance-prioritized feature distribution by incorporating objective factors during training. CoCaMIL achieved top classification performance across three large benchmarks and established a reliable system for grading sample difficulty. To our knowledge, CoCaMIL is the first approach to construct WSI morphological representations based on the collaborative integration of complexity factors, offering a new perspective to broaden the clinical use of deep learning in digital pathology. The code is available at https://github.com/sm8754/cocamil.