HIA-MTL: A dynamic hard task mining approach for breast and multi-organ histopathological image classification

Abstract

Breast cancer is the most common cancer among women worldwide, and its accurate classification based on digital pathology images is crucial for improving treatment precision. However, breast histopathology image classification faces challenges such as coarse-grained annotations and limited sample sizes, which restrict dataset expansion and pose significant difficulties for both manual examination and conventional deep learning methods. To address these issues, this paper proposes a Hard Instance Aware Meta-Transfer Learning (HIA-MTL) method for digital pathology images, aiming to tackle medical image analysis under few-shot, coarsely-labeled, and multi-class scenarios. The proposed method introduces a dynamic hard task and hard instance mining mechanism within a meta-learning framework. it integrates the traditional accuracy-based hard-task batching strategy by integrating instance-level hard-instance assessment with task-level dynamic hard-task mining.thereby more effectively utilizing limited annotated data. Experiments on the BreakHis dataset show that our method achieves classification accuracies of 99.71% and 97.73% for the 5-class and 8-class tasks, respectively, significantly outperforming current state-of-the-art methods. Furthermore, generalization experiments on several public medical image datasets, including BACH, SIPAKMED, and Kather-CRC-2016, demonstrate the promising cross-dataset generalization performance of the proposed approach on public benchmarks.