MLFI-Net: A Modified Encoder-Decoder Network With Multi-Layer Feature Interwoven for Liver Tumor Segmentation
Meiqin Chen, Xiaoliang Jiang, Weili Lu, Chunxian Peng · IEEE Access · 2025
Timely screening of liver tumors plays a crucial role in minimizing the risk of tumor deterioration and improving the survival rate of patients. However, the locations of liver tumors in different patients vary greatly, and their morphological characteristics often show significant differences. Moreover, the boundary between tumor tissue and the surrounding healthy liver parenchyma is not clear, which makes the accurate segmentation of liver tumors in CT images a significant challenge. To handle these limitations, we present a modified encoder-decoder network with multi-layer feature interwoven for liver tumor segmentation, named MLFI-Net. In our approach, we build upon the widely used encoder-decoder architecture as the foundational framework. To enhance feature extraction capabilities, we replace the standard convolutional modules with multi-scale atrous residual block (MARB). Additionally, we optimize the skip connection pathways by integrating multi-layer feature interwoven module (MFIM), which enriches the flow of multi-level contextual information between the encoder and decoder stages. In the decoding stage, we further refine the output by employing a multi-branch attention fusion block (MAFB), which selectively emphasizes informative features while suppressing irrelevant or redundant responses. Finally, we conducted a large number of experiments on our self-built dual-phase liver lesion datasets (arterial phase and portal venous phase) and the widely recognized 3DIRCADb dataset, which demonstrated the effectiveness of MLFI-Net in improving the accuracy of liver tumor segmentation. Specifically, on the arterial phase dataset, MLFI-Net attained Dice of 0.8404, Mcc of 0.8367, and Jaccard of 0.7264. Meanwhile, on the portal venous phase dataset, the results further improved, with Dice of 0.8447, Mcc of 0.8414, and Jaccard of 0.7360. When applied to the 3DIRCADb dataset, MLFI-Net exhibited excellent generalizability, achieving Dice of 0.8844, Mcc of 0.8848, and Jaccard of 0.7972. Furthermore, a series of ablation studies were conducted to evaluate the effectiveness of multi-scale atrous residual block, multi-layer feature interwoven module and multi-branch attention fusion block.