The article presents CaRe-CNN, a cascading refinement convolutional neural network (CNN) model for segmenting myocardial infarcts and microvascular obstructions in magnetic resonance imaging (MRI). The model is fully 3D and trained end-to-end in three stages to refine its predictions. It showed significant improvements in segmenting myocardial infarct tissue, a clinically relevant task. The model was ranked second in the FIMH 2023 MYOSAIQ challenge, outperforming in eight out of ten metrics. The accuracy of the model could help generate patient-specific heart models, contributing to personalized medicine.
Publication date: 19 Dec 2023
Project Page: https://orcid.org/0000-0002-6589-6560
Paper: https://arxiv.org/pdf/2312.11315
