Introduction of topologial features to advance medical imaging analysis

Medical Imaging is of pivotal importance for diagnosis and understanding of the diseases. Conventional automated medical imaging pipelines usually begin with the identification of a region of interest (ROI), followed by the extraction of a complete set of radiomic features, well characterizing the ROI texture and shape. Therefore, there is a lack of consideration of ROI morphological topology. To address this gap, we have developed a novel network-based feature set that complements radiomics. We introduce a novel approach to extract network-based features from 2D-3D binary ROIs, characterizing the topology. The starting point is the ROI’s skeleton extraction, from which it was possible to derive the graph representation. We evaluated the performances of the new feature set in four different contexts: femur landmark computation, brain aging prediction, endometriosis analysis and lung fibrosis analysis. The application showed promising results, allowing precise landmark identification and adding complementary information, increasing classification capability of the other tasks. The results demonstrate the potential of our framework to uncover new insights and enhance clinical outcomes. In summary, we have devised and tested a novel set of features that characterize the topology of ROIs in medical images expanding the array of available imaging features and improving understanding across various clinical scenarios.