Breast Tumor Image Classification in Bright Challenge VIA Multiple Instance Learning and Deep Transformers

Artificial intelligence models have become increasingly promising in automated computed-aided cancer diagnos-tics. In this paper, a new deep learning method is proposed for solving breast tumor image classification in the BRIGHT Challenge. Given two types of data, regions of interest (ROIs) and whole slide images (WSIs), the proposed method first uti-lize ROI data to train a model that is able to select important image patches in WSI data and simultaneously capture a low-dimensional representation for image patches. With features extracted from important image patches in each WSI, another deep learning model following the Transformer framework is trained to perform the final WSI-level tumor classification, in which transfer learning is also employed to fully exploit ROI data. Evaluated on the test dataset, the proposed method achieves the 4th best results in the challenge. Ablation exper-iments are also carried out to analyze the proposed method in detail.