An XAI approach to deep learning models in the detection of ductal carcinoma in situ

Abstract

Deep Learning models have been employed to improve the detection of medical problems using diverse imaging systems, following several challenges in which important models were created; especially related to breast cancer. However, their use has been limited in their application in the clinical domain around the world even though they improved the detection of breast cancer in women at an early stage. This provided the motivation for this work, whereby our contribution attempts to not only contribute to such an interesting field of study, but also to improve the early detection of breast cancer while enhancing clinicians’ confidence in such techniques through the practice of eXplainable Artificial Intelligence (XAI). The first part of the study followed a two-fold approach, taking-off with the collection of crucial information from the local specialist radiologists and surgeons in the field of breast cancer. The data collected through a questionnaire helped shed light about their practices, as well as their attitudes towards the employment and effectiveness of AI technologies. The similarity of the findings to the international literature encouraged us further to pursue our focus on addressing the lack of implementation of assistive AI models in hospitals, but also to investigate further how the introduction of XAI could further assist these professionals understand how a model classifies between a benign, malignant, or non-tumorous breast cancer thereby increasing their level of trust in such systems. This was followed by a comparative study between different available datasets to adopt as part of our scientific tests that employed a Convolutional Neural Network (CNN). The CBIS-DDSM scanned film mammography image dataset was selected due to its unique features, being that is has been carefully annotated by expert radiologists and has also been extensively used in the deep learning community. Additionally, we used this dataset on an already trained model, due to a lack of computational resources, and further focus on the element of XAI in deep learning. Our greatest contribution to the study was that of researching the best way to backpropagate the selected CNN model. This proves that it is possible to uncover the intricacies involved, at neuron level within a model, in converging towards a classification of a mammogram. After conducting several tests using different back-propagation methods, we noted that the Deep Taylor Decomposition and the LRP-Epsilon techniques produced the best results. These were obtained on a subset of 20 mammograms chosen at random from the CBISDDSM dataset. The results prove that XAI can indeed be used as a proof of concept to begin discussions on the implementation of assistive AI systems within the clinical community. The same experts who were interviewed in the initial stage of the study were further engaged in a post-study discussion around the interpretation of heatmaps generated by our research in which a critical feedback on the outcome was provided.