Multi-encoder U-Net benchmarking for LiTS17 Liver-Tumor segmentation: accuracy-efficiency trade-offs across training durations

Accurate liver and tumor segmentation from CT is fundamental for diagnosis, treatment planning, and longitudinal monitoring of liver cancer. Although U-Net variants with popular encoder backbones are widely used, the coupled effects of encoder selection, training duration, and computational cost, as well as comparisons against volumetric architectures such as V-Net, remain insufficiently standardized.

We propose a unified benchmarking framework that evaluates a family of multi-encoder 2D U-Net models together with an optional 3D V-Net baseline under the same preprocessing, input construction, and 3-fold cross-validation protocol on LiTS17. Multiple backbones (VGG16/VGG19/ResNet34/ResNet50/ResNet101/MobileNetV2) are assessed under 15/50/100-epoch schedules, and performance is reported using overlap and detection metrics (Dice, IoU, precision, recall) alongside efficiency indicators (training time and model complexity when available) to characterize the accuracy-efficiency trade-off.

Results show liver segmentation rapidly reaches near-ceiling performance across models, while tumor segmentation benefits markedly from longer training and stronger encoders, especially for small or low-contrast lesions.

Overall, the study provides a reproducible protocol and practical guidance for selecting segmentation models that balance accuracy, robustness, and deployment cost.