MEP-Net: a PIDNet-based model with median-enhanced spatial-channel attention for segmentation of hepatocellular...

Hepatocellular carcinoma (HCC) remains a major global health concern due to its high incidence and mortality. Contrast-enhanced ultrasound (CEUS) offers notable advantages for HCC diagnosis, including real-time imaging and non-invasiveness. However, CEUS images are often affected by noise, uneven signal distribution, and unclear boundaries between lesions and surrounding tissues, which pose challenges for automatic lesion segmentation.

To improve segmentation performance, this study proposes an improved PIDNet-based model termed MEP-Net. The model integrates a median-enhanced spatial-channel attention mechanism (MECS) and an efficient channel attention (ECA) module to enhance lesion-related feature representation and multi-branch feature fusion. We evaluate the model on a self-built CEUS dataset and the publicly available BUSI breast ultrasound dataset. It is also compared with several mainstream semantic segmentation methods, and ablation studies are conducted to analyze the contribution of each module.

The results show that MEP-Net outperforms the baseline PIDNet by 1.95%, 1.25%, and 2.51% in Dice, MIoU, and Recall, respectively, on the CEUS dataset, and by 1.37%, 1.06%, and 2.41% on the BUSI dataset. In addition, MEP-Net is compared with eight semantic segmentation methods and demonstrates superior overall segmentation performance and improved lesion representation. Ablation studies further confirm the complementary benefits of the MECS and ECA modules in improving segmentation accuracy.

The proposed MEP-Net achieves improved performance in CEUS image segmentation. By introducing attention mechanisms tailored to ultrasound image characteristics, it provides an effective approach for automatic HCC lesion segmentation.