Fast and Precise Liver Oncology Mapping: An IAROA-Optimised Dilated-Attention UNet Framework for Multitarget Abdominal CT Delineation

Liver oncology mapping involves the detailed visualization and analysis of liver tumors to assist in diagnosis, treatment planning, besides monitoring of liver cancer. This technique integrates imaging data (such as CT, MRI, or PET) with AI-driven segmentation and classification to identify tumor boundaries, stages, and vascular involvement. It enhances surgical planning, radiation targeting, and personalized therapy by providing a comprehensive spatial map of cancerous regions. Accurate delineation of hepatic structures and lesions in computed-tomography (CT) scans is critical for surgical planning, radio-frequency ablation, and intra-operative navigation. To present LD-DSC-SA, a lightweight yet high-precision segmentation framework that combines depth-separable convolutions (DSC), atrous-dilated receptive fields, and self-attention (SA) blocks, with its hyper-parameters automatically tuned by the improved artificial rabbit optimization algorithm (IAROA). The network was trained and evaluated on a harmonized corpus of 2 347 axial slices drawn from the 3DIRCADb01 and LiTS-17 datasets and tested on 393 unseen slices. LD-DSC-SA attains Dice scores of 95.23 % for liver parenchyma, 84.71 % for tumour core, and 80.37 % for hepatic veins—exceeding the conventional U-Net by 3.9 pp, 4.7 pp, and 3.2 pp respectively, while reducing parameters to 5.6 M and achieving 72 FPS inference on a single RTX 3080 GPU. An ablation study confirms that dilated convolutions and self-attention each contribute ~1–2 pp Dice gain, whereas focal loss and IAROA deliver the most significant uplift in class-imbalanced tumour segmentation. Statistical analysis using the Wilcoxon signed-rank test shows the performance improvements to be highly significant (p < 0.001) across all structures. Qualitative inspection further reveals crisper tumour boundaries and finer third-order vascular delineation compared with Attention U-Net and MobileNetV2-U-Net baselines. These consequences demonstrate that LD-DSC-SA offers an advantageous balance of accuracy, speed, and memory footprint, making it a viable candidate for real-time liver-surgery guidance and automated radiology workflows.