A Multi-Scale Micro-Structure Residual Enhancement Network for Liver Small Tumor Segmentation

This paper proposes a liver small tumor segmentation method based on a multi-scale micro-structure residual augmentation module. To address challenges such as small tumor size, blurred boundaries, and low contrast, improvements are made upon the BasicUNet architecture. The MSD Task03_Liver dataset is preprocessed using HU windowing and normalization, and three-dimensional CT images are converted into 512×512 slices for training. An encoder–decoder framework is adopted, where a spatial and channel Squeeze-and-Excitation (scSE) attention mechanism is embedded in skip connections to enhance feature selectivity. A multi-scale residual augmentation module with multi-branch depthwise separable convolutions is introduced in the decoder to strengthen fine-grained structural representation. Experimental results show that the proposed model significantly improves the Dice score and reduces the HD95 metric compared with the standard UNet baseline, achieving more accurate boundary segmentation and superior performance in small tumor delineation.