DYNAMIC DEFORMABLE ATTENTION AND GRADIENTENHANCED TRANSFORMERS FOR PRECISION 3D LIVER TUMOR ANALYTICS

Primary liver cancer, predominantly Hepatocellular Carcinoma (HCC) and Intrahepatic Cholangiocarcinoma (ICC), remains one of the leading causes of cancer-related mortality
worldwide. Accurate clinical management relies heavily on precise 3D volumetric lesion segmentation and non-invasive radiogenomic subtyping from multi-modal abdominal
Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) scans. However, current Deep Learning (DL) architectures encounter severe limitations due to extreme intratumoral heterogeneity, highly irregular lesion morphology, ill-defined boundaries, and severe
class imbalance. To resolve these challenges, this paper presents a novel, innovative hybrid
architecture termed D-LKA-UNETR++ (Deformable Large Kernel Attention UNETR++).
Our proposed model integrates three core algorithmic innovations: (1) a Deformable Large
Kernel Attention (D-LKA) encoder mechanism that expands receptive fields while
dynamically adapting to complex spatial lesion boundaries; (2) a Bottleneck GradientEnhanced Context Extraction (G-CE) module that preserves structural multi-frequency
features and prevents loss of spatial gradients; and (3) a Dual-Cross Attention (DCA) feature
fusion decoder that mitigates semantic gaps between encoder features and upsampled
representations. Benchmarking across public clinical repositories including the LiTS
benchmark dataset, 3D-IRCADb01, and MSD Task08 demonstrates that D-LKA-UNETR++
achieves state-of-the-art segmentation fidelity, yielding a Dice Similarity Coefficient (DSC)
of 0.912 ± 0.018 for tumor lesions and 0.978 ± 0.005 for organ parenchyma. Furthermore,
integrating multi-phase radiomic feature extraction with deep bottleneck embeddings enables