Comment on "Automatic rib fracture detection on postmortem CT data using deep learning".

Lopez-Melia et al. recently reported a feasibility study on automatic rib fracture detection in postmortem computed tomography (PMCT) using nnDetection, a self-configuring three-dimensional object detection framework based on deep learning (DL), and explored the impact of domain shift between clinical and postmortem datasets. This comment discusses three issues that may refine the interpretation of their findings and inform future work. First, we argue that the relationship to previous PMCT DL studies by Ibanez et al., including RiFNet and a subsequent fracture-type classifier, should be clarified, and that the term “detection” needs to be explicitly defined in contrast to rib-wise or patch-based classification. Second, we highlight potential circularity in the reference standard, because the same radiologist created and revised the annotations partly guided by the PMCT-trained nnDetPM output, and we suggest a multi-reader, partially blinded consensus approach for future studies. Third, we discuss the domain shift experiments between clinical and PMCT imaging, including the case selection within the postmortem rib fracture (PMRF) dataset, the use of a low intersection-over-union (IoU) threshold, and the implications for generalizability and cross-domain model development. We conclude that the study is an important step toward practical computer-aided detection of rib fractures on PMCT, but that clearer positioning relative to prior work, stronger ground-truth procedures, and explicit consideration of harmonization and domain adaptation would further enhance its impact for forensic practice.