Trans-Attention Multiple Instance Learning for Cancer Tissue Classification in Digital Histopathology Images

Abstract

The detection of cancerous tissue in histopathological slides is of great value in both clinical practice and pathology research. This paper presents a novel approach that targets automatically classifying cancer tissue by leveraging an attention multiple instance learning scheme; an attention-equivalent neural network-based permutation-invariant aggregation operator applied on the multi-instance learning network. Additionally, we propose a Trans-AMIL approach which is designed to apply Transfer Learning pre-trained models and learn the distribution of the bag label probability using neural networks. We demonstrate experimentally that our approach outperforms several conventional deep learning-based methods on an open BreakHis cancer histopathology dataset.