Combining Deep Learning and ASP-Based Models for the Semantic Segmentation of Medical Images

Automatic segmentation represents a huge breakthrough in computer-aided diagnosis and medicine, as it allows to provide clinicians important with information for interventional and diagnostic tasks. Recent advancements in Deep Learning (DL), such as Convolutional Neural Networks (CNNs), have proved to be greatly promising in identifying anatomical and pathological structures, and in extracting meaningful patterns from huge amounts of data. However, such approaches suffer from the lack of proper means for interpreting the choices made by the models, and it is not easy to drive the decisions according to prior knowledge. In this context, deductive rule-based approaches, such as Answer Set Programming (ASP), can allow to effectively encode problems or specific features via logic programs in a declarative fashion, while possibly also helping at improving performance. In this seminal work, we propose the use of ASP to drive DL approaches in performing semantic segmentation of medical images. Specifically, we encoded prior medical knowledge via ASP, thus defining a rule-based model for deducting all admitted combinations of classes and right locations in medical images. The results of an experimental analysis are reported with the aim to assess the viability of the proposed approach.