Unlocking Clinical Potential: Beyond Single-to-Tri-Phase CT with Dynamic Fusion for Precision Liver Tumor Segmentation

JinYue Li, Qiankun Li, Huabao Chen, Feng He, Kun Wang, Xin Ning, Prayag Tiwari

Keywords: Liver tumor segmentation, contrast-enhanced CT, multi-phase deep fusion, phase-wise attention mechanism

TL;DR: We introduce a clinically guided multiphase fusion framework that unlocks radiological potential for AI, leveraging the largest and most comprehensive liver CECT dataset to achieve significant performance gains.

Abstract: Liver tumor segmentation is essential for treatment planning and disease monitoring. Most existing methods rely on single-phase computed tomography (CT), they often suffer from low contrast and incomplete lesion depiction. Contrast-Enhanced CT (CECT) offers multiple imaging phases: arterial (ART), portal venous (PV), and delayed (DL), which provide complementary anatomical and functional information. This study begins with a systematic quantitative evaluation of each enhanced phase using standard segmentation models to investigate their individual contributions and validate phase-specific clinical insights. Guided by this analysis, a Multi-phase Attention Deep Fusion Network (MADF-Net) is proposed to hierarchically integrate ART, PV, and DL features across the input, feature, and decision levels. Experiments on the clinically collected multi-phase liver lesion (MPLL) dataset (the largest and most clinically comprehensive multi-phase liver cancer CECT dataset) demonstrate that the proposed method achieves state-of-the-art segmentation performance. MADF-Net achieves a Dice score of 78.65%, which is 9.39 higher than the best single-stage baseline, by deeply fusing information from three phases, and consistently improves across all evaluation metrics. Our codes are available at: https://anonymous.4open.science/r/ICLR26_unlocking_clinical_potential-EFE8/.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 8668