A Parallelization Approach for Hybrid-AI-based Models: an Application Study for Semantic Segmentation of Medical Images

In the last decades, Artificial Intelligence (AI) approaches have been fruitfully employed in many tasks; for instance, Deep Learning (DL)-based methods have shown great ability in extracting meaningful features from images, providing valuable support to computer-aided diagnosis and medicine. Including prior knowledge in DL-based approaches could help in making their decisions more powerful, understandable, and explainable. However, even if this combination has raised a lot of interest in the scientific community, still remains an open problem due to several difficulties, for example, in modeling complex domains, handling missing specifications, and identifying the most suitable architecture able to properly combine the two AI worlds. In this work, we rely on an existing framework defined for combining deductive and inductive approaches; in particular, explicit knowledge is encoded using Answer Set Programming (ASP), included in the training, and used to improve the quality of the images via a post-processing phase. We propose a parallelization of this approach that drastically reduces the execution time. The proposed approach has been tested using different neural networks for semantic segmentation tasks over Laryngeal Endoscopic Images.