LiverSegNeXt: Improved MedNeXt for Liver Segmentation

Liver segmentation is crucial for computer-assisted diagnosis and preoperative planning in hepatology. The scarcity of annotated medical datasets and segmentation architectures tailored for liver imaging, presents significant challenges. Inaccurate segmentation can severely impact the diagnosis of intrahepatic lesions and preoperative planning. Current state-of-the-art methods struggle with segmenting challenging liver regions, such as the hepatic apex, portal vein, base, and caudate lobe. To address this, we introduce LiverSegNeXt, an advanced technique based on the MedNeXt framework. LiverSegNeXt incorporates novel approaches, including the inception and ConvNeXt architecture. Additionally, the performance of existing methods is still limited by insufficient feature extraction. We propose a coarse training and fine-tuning approach, as well as an absolute position embedding approach, to leverage the consistent spatial positioning of the liver in CT scans. We evaluated LiverSegNeXt on our in-house dataset. Our method outperforms existing models and found that it surpasses other existing methods, improving the Dice coefficient by 0.21% compared to other approaches.