Semi-Supervised Histology Classification using Deep Multiple Instance\n Learning and Contrastive Predictive Coding

Abstract

Convolutional neural networks can be trained to perform histology slide\nclassification using weak annotations with multiple instance learning (MIL).\nHowever, given the paucity of labeled histology data, direct application of MIL\ncan easily suffer from overfitting and the network is unable to learn rich\nfeature representations due to the weak supervisory signal. We propose to\novercome such limitations with a two-stage semi-supervised approach that\ncombines the power of data-efficient self-supervised feature learning via\ncontrastive predictive coding (CPC) and the interpretability and flexibility of\nregularized attention-based MIL. We apply our two-stage CPC + MIL\nsemi-supervised pipeline to the binary classification of breast cancer\nhistology images. Across five random splits, we report state-of-the-art\nperformance with a mean validation accuracy of 95% and an area under the ROC\ncurve of 0.968. We further evaluate the quality of features learned via CPC\nrelative to simple transfer learning and show that strong classification\nperformance using CPC features can be efficiently leveraged under the MIL\nframework even with the feature encoder frozen.\n

Ähnliche Arbeiten