Deep Learning-Based Liver and Tumor Segmentation: A Comparative Analysis on the LiTS Dataset

Abstract

The segmentation of liver and liver tumors plays a critical role in clinical oncology, radiology, and surgical applications. Manual or semi-automatic segmentation methods are not only time-consuming, but also prone to subjectivity. In contrast, deep learning-based methods offer an automated approach that delivers objective and reproducible results. In this study, we evaluate several deep learning architectures—including established models such as U-Net and ResNet-based methods—and introduce novel segmentation algorithms that have been successfully applied in other imaging contexts yet remain untested on liver images. The Liver Tumor Segmentation Benchmark (LiTS) dataset, with its large number of annotated liver and tumor images, has been selected to ensure a fair and comprehensive evaluation. Segmentation performance is assessed using supervised metrics such as precision, recall, Intersection over Union (IoU), and Dice coefficient. Experimental results suggest that the optimal segmentation method can significantly aid physicians in accurate diagnosis and treatment planning, while also serving as a valuable tool in medical education and clinical decision support systems. This innovative approach not only advances the field of medical image analysis, but also has the potential to enhance clinical workflows and enable personalized treatment strategies. This study compares the performance of various deep learning-based methods for liver and tumor segmentation. Using the LiTS dataset, we evaluated the performance of U-Net, DeepLabV3+, and SegFormer models. IoU and Dice coefficients were utilized to measure segmentation success. The U- Net model demonstrated high performance with limited data, while DeepLabV3+ yielded better results on large datasets. SegFormer showed promising performance, highlighting the effectiveness of the Vision Transformer-based approach.