Compare three modeling paradigms for predicting liver resection zones, with attention to anatomical and functional inputs relevant to parenchyma-sparing planning.
Approach:
Dataset and preprocessing: The retrospective dataset included 104 patients and 104 reference resection zones. Segmentation masks were rasterized onto a shared 0.4 mm isotropic voxel grid.
Distance-based model: Used liver boundaries and tumor location to model a non-anatomical resection for surface-near tumors. The tumor was dilated by 8 mm, based on the dataset’s mean minimal resection margin.
Perfusion-based model: Used liver, tumor, and vascular structures. Portal and hepatic vein trees were represented as directed graphs to estimate downstream perfusion loss after vessel transections.
Deep learning model: The article context identifies deep learning as the third modeling paradigm but does not provide its methods.
Evaluation: Predicted and reference zones were to be compared using overlap measures (DSC and IoU) and distance measures (95th-percentile and maximum Hausdorff distances, and average surface distance).
Key Findings:
The retrospective cohort’s mean planned in silico margin was 7.63 ± 6.70 mm, approximately 8 mm.
The supplied context describes the study design and methods but does not report comparative model performance results.
Interpretation:
The study compares distance-based, perfusion-based, and deep learning approaches for resection-zone prediction; the available context does not establish which approach performed best.
Limitations:
The provided article context ends during the perfusion-based model description and omits the deep learning methods and results.
The analysis used retrospective data from 104 patients at one medical center.
