Innovations in Medical Image Analysis and Explainable AI for Transparent Clinical Decision Support Systems

This thesis explores innovative methods designed to assist clinicians in their everyday practice, with a particular emphasis on Medical Image Analysis and Explainability issues. The main challenge lies in interpreting the knowledge gained from machine learning algorithms, also called black-boxes, to provide transparent clinical decision support systems for real integration into clinical practice. For this reason, all work aims to exploit Explainable AI techniques to study and interpret the trained models. Given the countless open problems for the development of clinical decision support systems, the project includes the analysis of various data and pathologies. The main works are focused on the most threatening disease afflicting the female population: Breast Cancer. The works aim to diagnose and classify breast cancer through medical images by taking advantage of a first-level examination such as Mammography screening, Ultrasound images, and a more advanced examination such as MRI. Papers on Breast Cancer and Microcalcification Classification demonstrated the potential of shallow learning algorithms in terms of explainability and accuracy when intelligible radiomic features are used. Conversely, the union of deep learning and Explainable AI methods showed impressive results for Breast Cancer Detection. The local explanations provided via saliency maps were critical for model introspection, as well as increasing performance. To increase trust in these systems and aspire to their real use, a multi-level explanation was proposed. Three main stakeholders who need transparent models have been identified: developers, physicians, and patients. For this reason, guided by the enormous impact of COVID-19 in the world population, a fully Explainable machine learning model was proposed for COVID-19 Prognosis prediction exploiting the proposed multi-level explanation. It is assumed that such a system primarily requires two components: 1) inherently explainable inputs such as clinical, laboratory, and radiomic features; 2) Explainable methods capable of explaining globally and locally the trained model. The union of these two requirements allows the developer to detect any model bias, the doctor to verify the model findings with clinical evidence, and justify decisions to patients. These results were also confirmed for the study of coronary artery disease. In particular machine learning algorithms are trained using intelligible clinical and radiomic features extracted from pericoronaric adipose tissue to assess the condition of coronary arteries. Eventually, some important national and international collaborations led to the analysis of data for the development of predictive models for some neurological disorders. In particular, the predictivity of handwriting features for the prediction of depressed patients was explored. Using the training of neural networks constrained by first-order logic, it was possible to provide high-performance and explainable models, going beyond the trade-off between explainability and accuracy.