Research on automatic segmentation of hepatocellular carcinoma in CT images based on an improved nnU-Net

Abstract

Accurate 3D segmentation of hepatocellular carcinoma (HCC) is crucial for preoperative planning; however, extreme scale variations, inherently blurred boundaries, and the prohibitive computational overhead associated with 3D convolutions pose formidable challenges. To address this, we propose DualConvNeXt-U, a novel 3D segmentation architecture designed to achieve an optimal balance between high precision and lightweight efficiency. Built upon the nnU-Netv2 framework, the network introduces a macro-level design incorporating ConvNeXt-style 7×7×7 depth-wise separable large kernels and inverted bottlenecks to effectively expand the receptive field. At the micro-level, it embeds a DualConv grouped dual-kernel mechanism (comprising 3×3×3 and 1×1×1 convolutions), which efficiently mitigates cross-channel redundancy at a minimal computational cost while refining multi-scale edge representations. Extensive experiments on the multi-center LiTS 2017 dataset demonstrate that our model significantly outperforms state-of-the-art (SOTA) methods, achieving a Dice coefficient of 77.15% and a 95% Hausdorff Distance (HD95) of 7.25 mm. Furthermore, the parameter count (14.6 M) and computational complexity (128.2 G FLOPs) are reduced by over 60% compared to the baseline, highlighting its remarkable potential for clinical deployment on edge devices.