Efficient Liver and Tumor Segmentation Using a Compact Residual Network and Contrast-Enhanced Pre-Processing Pipeline

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate segmentation of liver and tumors in CT images is crucial for clinical decision-making.
• Existing methods may lack efficiency or generalizability across different datasets.

Purpose of the Study:

• To develop and evaluate SegResNet_2335, a lightweight 3D residual network for volumetric segmentation of liver and tumors in CT images.
• To assess the framework's performance and cross-dataset generalization capabilities.

Main Methods:

• A tailored pre-processing pipeline including voxel spacing resampling, CT window adjustment, CLAHE, and z-score normalization was applied.
• A lightweight 3D residual network (SegResNet_2335) with 1.5 million parameters was utilized for segmentation.
• The framework was evaluated on the LiTS and 3D-IRCADb-01 datasets.

Main Results:

• The model achieved high Dice Similarity Coefficients (DSCs): 0.956 for liver and 0.754 for tumor on the LiTS test set.
• Cross-dataset evaluation on 3D-IRCADb-01 yielded a liver DSC of 0.847 and a tumor DSC of 0.706.
• Rapid inference time of approximately 1.8 seconds per scan was demonstrated.

Conclusions:

• SegResNet_2335 offers strong and consistent liver and tumor segmentation performance.
• The framework exhibits robust cross-dataset generalization and is suitable for resource-constrained clinical deployment.
• The publicly available implementation promotes reproducibility in medical image analysis research.