Dynamic Swin-UNet with Adaptive Fusion for Precision HCC Delineation and Early Diagnosis

Hepatocellular Carcinoma (HCC) and primary hepatic malignancies rank among the leading causes of
cancer-related mortality globally. Accurate, early-stage diagnosis and automatic lesion segmentation from
dynamic multiphase computed tomography (CT) and magnetic resonance imaging (MRI) are paramount
for surgical planning, targeted local therapy, and treatment efficacy monitoring. However, computerized
liver tumor segmentation remains an intricate challenge due to low tissue contrast, heterogeneous lesion
morphology, ill-defined boundaries, microvascular invasion, and significant intra-organ structural
variability. To address these critical clinical gaps, this paper presents a novel, highly innovative deep
learning architecture named Multi-Phase Dynamic Swin-UNet Transformer with Adaptive Contrastive
Fusion (MPD-SwinUNet-ACF). The proposed framework integrates hierarchical multi-scale Shifted
Window (Swin) Transformer blocks within an encoder-decoder backbone to capture comprehensive longrange contextual spatial dependencies without incurring prohibitive computational complexity.
Furthermore, an Adaptive Contrastive Dynamic Fusion (ACDF) module is engineered to synergistically fuse
feature representations across arterial, portal venous, and delayed imaging phases, capturing microvascular dynamics essential for differential diagnosis. To refine fuzzy, ambiguous tumor margins, a
specialized Boundary-Guided Gradient Refinement (BGGR) block incorporates high-frequency Sobel and
Laplacian feature maps directly into the skip connections. Extensive experimental validations conducted
on benchmark datasets—including the Liver Tumor Segmentation Challenge (LiTS 2017), 3DIRCADb, and
an extended 2025–2026 multi-center clinical MRI/CT cohort—demonstrate that MPD-SwinUNet-ACF
achieves state-of-the-art performance with a Dice Similarity Coefficient (DSC) of 97.42% for liver
parenchyma and 92.18% for hepatic tumor segmentation, outperforming conventional U-Net, ResUNet++,
MedNeXt, and hybrid CNN-Transformer baselines. Computational efficiency evaluations reveal an average
inference latency of 41.2 milliseconds per 3D volume, verifying its suitability for real-time intraoperative
navigation and clinical decision support.