ComboLoss for NECT Liver Tumor Segmentation: An Imbalance-Aware nnU-Net v2 Baseline

Non-enhanced CT (NECT) liver tumor segmentation remains challenging due to weak lesion–parenchyma contrast, severe foreground–background imbalance, and the prevalence of very small lesions. To provide a compact and reproducible baseline under small-data conditions, we build on nnU-Net v2 (3D full-resolution) and introduce ComboLoss, a weighted combination of Dice loss and focal loss that increases emphasis on hard positive voxels while remaining fully compatible with nnU-Net training mechanics such as deep supervision and ignore-label masking. Experiments are conducted on an institute-provided NECT dataset consisting of 40 volumes with a fixed split of 28 training and 12 test cases. Performance is evaluated using overlap metrics (DSC, IoU, precision, recall), a boundary metric (HD95), and a detection-oriented lesion-level recall metric where a ground-truth lesion is considered detected if any predicted component overlaps it with IoU ≥ 0.1. Compared with the default Dice+CE objective, ComboLoss improves segmentation overlap and lesion detectability, achieving DSC 0.624 ± 0.365 and lesion-level recall@0.1 of 0.694 ± 0.440 on the test set. However, the results also show that overlap improvement may be accompanied by larger HD95 values when isolated false-positive components are produced. To address this issue, we further evaluate connected-component post-processing, which reduces boundary outliers while preserving comparable overlap and lesion-level detection performance. We also compare imbalance-aware objectives including Tversky Loss and Focal Tversky Loss under the same nnU-Net v2 configuration, providing a controlled analysis of loss design under severe class imbalance. The proposed setup offers a transparent and reproducible reference pipeline for NECT liver tumor segmentation in small-data scenarios.