Developing effective methods for Invasive Ductal Carcinoma (IDC) detection remains a challenging problem for breast cancer diagnosis. Recently, there have been notable success of utilizing deep neural networks in various application domains, however, it is well-known that training deep neural networks requires a huge amount of labeled training data in order to achieve high accuracy. Such amounts of manually labeled data are time-consuming and very expensive to obtain especially when domain expertise is required. In order to take advantage of cheap available unlabeled data, we present a novel semi-supervised learning framework for IDC detection using small amounts of labeled training examples. In order to gain trust in the prediction of the proposed framework, we explain the prediction globally. Our proposed framework consists of five main stages including data augmentation, feature selection, dividing co-training data labeling, deep neural network modeling and interpretability of the 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. The experimental evaluation shows that our framework is able to detect IDC with a balanced accuracy of 0.865 and F-measure of 0.773 using only 10% labeled instances from the training dataset while the rest of the training dataset is treated as unlabeled.