Abstract
The early detection of Hepatocellular Carcinoma (HCC) using computed tomography (CT) is impeded by high annotation costs, lesion heterogeneity, and inadequate anatomical topology modeling. This study proposes a Dual-Branch Collaborative Generative Adversarial Network (DB-Collab GAN) with Anatomical Topology Coding to address these challenges. The framework features a dual-branch architecture that forms a "segmentation-guided detection" loop, with cross-layer feature sharing enhancing local-global complementarity. A layered Multi-Scale Convolutional Block Attention Module (CBAM) captures micro-details via 1 × 1 convolutions and liver anatomy via 5 × 5 convolutions. Anatomically tailored sine-cosine coding embeds the Couinaud segment topology, reducing the mean localization error (ADE) to 3.01 mm. Semi-supervised adversarial optimization with a dual-path discriminator achieved performance comparable to 7,140 supervised cases using only 1,070 labeled cases. On 7,140 clinical CT slices, the method outperformed the baselines in terms of accuracy (0.8875 ± 0.02), recall (0.8613 ± 0.03), and F1-score (0.8848 ± 0.02), with a 10.66% higher F1-score than Mask RCNN. Ablation studies confirmed the contributions of the multiscale CBAM and topology coding. It maintains robustness under high noise (ADE = 4.57 mm), providing a low-annotation-dependent solution, effectively reducing missed diagnoses and misclassifications of small lesions and vascular artifacts, and supporting clinical decision-making in early intervention.