An interpretable semi-supervised framework for patch-based classification of breast cancer.

Developing effective invasive Ductal Carcinoma (IDC) detection methods remains a challenging problem for breast cancer diagnosis. Recently, there has been notable success in utilizing deep neural networks in various application domains; however, it is well-known that deep neural networks require a large amount of labelled training data to achieve high accuracy. Such amounts of manually labelled data are time-consuming and expensive, especially when domain expertise is required. To this end, we present a novel semi-supervised learning framework for IDC detection using small amounts of labelled training examples to take advantage of cheap available unlabeled data. To gain trust in the prediction of the framework, we explain the prediction globally. Our proposed framework consists of five main stages: data augmentation, feature selection, dividing co-training data labelling, deep neural network modelling, and the interpretability of neural network prediction. The data cohort used in this study contains digitized BCa histopathology slides from 162 women with IDC at the Hospital of the University of Pennsylvania and the Cancer Institute of New Jersey. To evaluate the effectiveness of the deep neural network model used by the proposed approach, we compare it to different state-of-the-art network architectures; AlexNet and a shallow VGG network trained only on the labelled data. The results show that the deep neural network used in our proposed approach outperforms the state-of-the-art techniques achieving balanced accuracy of 0.73 and F-measure of 0.843. In addition, we compare the performance of the proposed semi-supervised approach to state-of-the-art semi-supervised DCGAN technique and self-learning technique. The experimental evaluation shows that our framework outperforms both semi-supervised techniques and detects IDC with an accuracy of 85.75%, a balanced accuracy of 0.865, and an F-measure of 0.773 using only 10% labelled instances from the training dataset while the rest of the training dataset is treated as unlabeled.