UniRepLKNet-AHLFA: A Hybrid UniRepLKNet Framework with Adaptive Hierarchical Lesion Feature Aggregation for Liver Tumor CT Segmentation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Liver tumor segmentation from Computed Tomography (CT) images plays an integral role in computeraided diagnosis, treatment planning, and disease monitoring.However, existing liver tumor segmentation methods often suffer from high computational complexity, inadequate multi-scale feature representation, dependence on multistage segmentation pipelines, and reduced accuracy in delineating small and irregular tumor regions.To address these challenges, this study proposes UniRepLKNet-AHLFA, a hybrid deep learning (DL) framework that integrates the UniRepLKNet backbone with a proposed Adaptive Hierarchical Lesion Feature Aggregation (AHLFA) module.The proposed framework employs large-kernel hierarchical feature extraction, adaptive feature weighting, hierarchical multi-scale feature fusion, and context-aware feature refinement to enhance lesion-specific feature representation and improve liver tumor segmentation accuracy.The framework was trained and evaluated using the LiTS PNG dataset, comprising abdominal CT volumes converted into 256 × 256 two-dimensional images with corresponding groundtruth tumor masks.A CT-volume-wise partitioning strategy was employed to ensure that slices originating from the same CT volume were not distributed across the training, validation, and testing subsets, thereby preventing inter-slice information leakage across the development and evaluation partitions.In the primary single-run evaluation on the independent volume-wise test set, the proposed framework attained a Dice Similarity Coefficient (DSC) of 96.70%, an Intersection over Union (IoU) of 93.62%, 98.81% accuracy, 97.05% precision, 96.36% recall, 99.35% specificity, and an F1-score of 96.70%.Furthermore, volume-wise ablation analysis, group-wise five-fold cross-validation, and five-run controlled comparison with recent segmentation baselines demonstrated consistent and robust segmentation performance under leakage-controlled evaluation conditions.Computational profiling further quantified the model complexity and inference requirements, demonstrating that the architectural improvements were achieved with moderate computational overhead.These findings demonstrate that the proposed UniRepLKNet-AHLFA framework provides an effective approach for accurate liver tumor segmentation from CT images and exhibits reliable generalization to previously unseen CT volumes under the adopted evaluation protocol.