YOLOv8-Based Frameworks for Liver and Tumor Segmentation Task on LiTS

Accurate liver tumor diagnosis in clinical practice relies on precisely delineating the liver and identifying potential tumors in Computed Tomography scans. This study aims to develop a lightweight liver and tumor segmentation model, balancing computational cost and accuracy. Traditional methods use Convolutional Neural Networks that enhance the performance but heighten the computational costs. Addressing it, this study presents three frameworks: Sim-Seg, Seg-Seg, and BB-Seg for liver and liver tumor segmentation tasks, leveraging the You Look Only Once network. The proposed Sim-Seg framework achieves a dice score of 89.54% for liver segmentation, while the BB-Seg framework attains a dice score of 80.55% for tumor segmentation on the Liver Tumor Segmentation Benchmark. In conclusion, this study effectively balances computational resources and accuracy, achieving results comparable to those of more compute-intensive methods. It aims to advance the development of YOLO-based frameworks in the medical field.