BCIW-Net: Boundary-Conditioned Adaptive CT Windowing for Hepatocellular Carcinoma Segmentation.
Authors
Affiliations (5)
Affiliations (5)
- Department of Computer Engineering and Mathematics, Universitat Rovira i Virgili, 43007 Tarragona, Spain.
- Centre for Cancer Biomarkers & Biotherapeutics, Barts Cancer Institute, Queen Mary University of London, London EC1M 6BQ, UK.
- Department of Pathology and Clinical Bioinformatics, Erasmus MC, 3015 GD Rotterdam, The Netherlands.
- Queen's University Belfast, Belfast BT7 1NN, UK.
- Electrical Engineering Department, Faculty of Engineering, Aswan University, Aswan 81528, Egypt.
Abstract
<b>Background/Objectives:</b> Hepatocellular carcinoma (HCC) is the most common primary liver malignancy, and accurate tumour segmentation in contrast-enhanced computed tomography is essential for treatment planning, assessment of ablation margins, and post-treatment evaluation. Existing deep learning methods typically use fixed-intensity CT preprocessing, overlooking the adaptive windowing employed by radiologists to enhance tumour visibility. To address this limitation, we propose BCIW-Net, a boundary-conditioned intensity-windowed network that integrates spatially adaptive and differentiable intensity windowing into the segmentation process as a clinically motivated inductive bias. <b>Methods:</b> The proposed framework first generates complementary soft-tissue and tumour-enhancing intensity representations using differentiable sigmoid-based transformations. A primary encoder-decoder module predicts tumour guidance, boundary probabilities, and pixel-wise error proxy maps, which collectively guide the dynamic estimation of spatially varying window centre and width parameters. The resulting adaptively windowed features are selectively blended with fixed-intensity representations through an error-gated mechanism before final refinement in a secondary encoder-decoder network. A multi-task objective jointly optimises tumour segmentation, boundary prediction, error-proxy estimation, and adaptive window regularisation in an end-to-end manner. <b>Results:</b> Experiments on the primary liver cancer (PLC)-CECT dataset demonstrate that BCIW-Net achieved a mean Dice similarity coefficient of 78.68% and an intersection over union (IoU) of 64.85% across five-fold patient-level cross-validation, outperforming the seven evaluated CNN-, Transformer-, and state-space baselines. Evaluation on LiTS2017 without dataset-specific fine-tuning yielded a tumour Dice score of 89.84%, providing preliminary evidence of cross-dataset generalisation. <b>Conclusions:</b> Qualitative feature visualisations show that adaptive windowing enhances tumour representation and boundary delineation, demonstrating the effectiveness and interpretability of BCIW-Net for HCC segmentation.