A Cascaded Deep Learning Framework for Robust Liver CT Segmentation Using ROI Refinement and Patient-Level Cross-Validation

Abstract

Accurate liver segmentation from computed tomography (CT) images is essential for computer-aided diagnosis, treatment planning, volumetric assessment, and longitudinal monitoring of patients with hepatocellular carcinoma. However, automated segmentation remains challenging because of variations in liver morphology, low-contrast boundaries, pathological abnormalities, and neighbouring abdominal structures with similar intensity characteristics. This study presents a failure-aware cascaded deep learning framework for automated liver CT segmentation using the publicly available HCC-TACE-Seg dataset. The proposed approach first performs coarse liver localisation on the complete CT image and then extracts a region of interest with an additional safety margin. The cropped region is subsequently processed by a high-resolution U-Net++ model with a ResNet34 encoder. A hybrid Dice and binary cross-entropy loss is used during training, while validation-optimised thresholding, largest-component filtering, and morphological hole filling are applied to improve anatomical consistency. Monte Carlo dropout is further incorporated to estimate predictive uncertainty and support case-level quality assessment. Following annotation verification, 103 liver-positive representative CT images were retained after excluding one image with an empty liver mask. Five-fold cross-validation was performed at the available patient-identifier level to prevent samples from the same patient from being distributed across training and testing partitions. The proposed framework achieved a mean Dice similarity coefficient of 0.8171, Intersection over Union of 0.7442, precision of 0.7993, sensitivity of 0.8737, and specificity of 0.9750. Boundary-based evaluation yielded an HD95 of 48.91 pixels, an average symmetric surface distance of 13.41 pixels, and a surface Dice score of 0.4250. Probability calibration produced a Brier score of 0.0418 and an expected calibration error of 0.0811. Uncertainty estimates were additionally used to flag potentially unreliable segmentations for manual review. These findings indicate that cascaded localisation and region-of-interest refinement can provide robust liver segmentation while reducing background interference and supporting uncertainty-informed quality control.