Morphological Information Extraction in medical imaging using deep learning interpretability: application case on craniofacial dysmorphism

摘要: Recent advances in the interpretability of convolutional neural networks (CNNs) have allowed applications in imaging as a novel method for visual feature extraction. We used this approach to investigate the impact of changing occlusal forces on craniofacial architecture in class II retrognathism (C2Rm) pathology. Better understanding the points of impact of C2Rm on the entire skull is a major challenge in the diagnosis, treatment, and management of this dysmorphism, but also allows to be part of the debate on changes in the shape of the skull during human evolution. To address these challenges, we introduced a methodology for extracting morphological information with an interpretable CNN (MIE-ICNN) that combines: 1) a pre-processing step to train a deep CNN, 2) an interpretability step using the Score-CAM method to explain the areas of the radiographs used by the model to identify the pathological class, and 3) a calculation step to generate a global activation map, which represents an average activation map of all classifications. The main conclusions of this study are as follows: I) the proposed technique made it possible to find the anatomical areas affected by C2Rm and already identified in the literature (ie the cranial base and the vertebrae), to confirm the involvement of a controversial area (the frontal sinus), and to identify a new structure (the parietal bone) as a biomarker. In addition, the study of the involvement of anatomical zones according to the severity of the dysmorphism (by the same methodology), has made it possible to propose hypotheses on the evolution of the forms of the skulls during human evolution (highlighting, in …