Abstract The analysis of the liver, including its vessels and lesions, is essential for pre-operative intervention planning. Automated deep learning-based segmentation provides a time-efficient method for consistently localizing and visualizing these structures. However, the integration of deep learning algorithms into clinical practice necessitates thorough testing and validation. It is crucial not only to assess quantitative and volumetric metrics but also to evaluate these algorithms clinically through experienced radiologists and surgeons. In this study, we used an independent test set of 107 computed tomography scans to conduct a multi-user, multi-disciplinary analysis of a deep learning-based segmentation results. Six experienced clinicians-three surgeons, one surgical resident and two radiologists-evaluated the predictions made by the deep learning models. While liver segmentation can be considered largely solved, vessel segmentation still requires improvement. By manually identifying missing lesions (false negatives) as well as accepting and rejecting proposed lesions, performance metrics can be calculated for the lesion segmentation without needing a reference segmentation. Notably, lesion segmentation achieved an F1-score of 0.82 for lesions larger than 1 cm in diameter.
Kock et al. (Mon,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: