A Multi-Task Multiple Instance Learning Algorithm to Analyze Large Whole Slide Images from Bright Challenge 2022

Abstract

Malignant lesions in breast tissue specimen whole slide images (WSIs), may lead to a dangerous diagnosis, such as cancer. However, WSIs analysis is time-consuming and expensive, requiring the work of expert pathologists. This paper aims to present a method for the 2022 BRIGHT Challenge, that involves the analysis of breast WSIs. The organizers provided over 550 breast WSIs and over 3900 regions of interest (ROIs) to develop and validate methods for breast cancer images. The method presented in this work is based on a Multiple Instance Learning instance-based Convolutional Neural Network (CNN), allowing the combination of strongly-annotated data (from ROIs) and weakly-annotated data (from WSIs) via the optimization of a multi-task loss function. Furthermore, during the CNN training, the input patches are clustered and filtered according to their entropy, to reduce the non-informative content used to train the model. The CNN reaches an averaged <tex>$\text{F1-score}=0.63\pm 0.02$</tex> on the 3-class classification task and averaged <tex>$\text{F1-score}=0.39 \pm 0.08$</tex> on the 6-class classification task, considering the val-idation partition; an averaged <tex>$\text{F1-score}=0.65$</tex> on the cancer risk classification task and averaged <tex>$\text{F1-score}=0.45$</tex> on the sub-typing cancer risk classification task, considering the best result achieved on the test partition. These results show that Multiple Instance Learning instance-based CNNs may represent a good resource to tackle this kind of problem.