Multimodal Liver Tumor Image Segmentation Model Based on Improved YOLO

To address the clinical need for real-time liver tumor segmentation and to overcome the two main limitations of the U-Net model's high computational cost and the YOLO framework's inadequate local feature modeling, this paper proposes a YOLO-Med model based on an improved YOLOv8.In terms of image preprocessing, the model dynamically optimizes input CT and MRI images by introducing a joint adjustment mechanism based on Contrast-to-Noise Ratio (CNR) and Signal-to-Noise Ratio (SNR). For the backbone network, a residual attention module is introduced to dynamically enhance tumor region responses. When combined with a multi-scale mask branch, the model achieves progressive refinement from contour to boundary, improving tumor segmentation accuracy and enhancing boundary delineation. Additionally, to address modality differences between CT and MRI images, a feature alignment network and a cross-modality fusion segmentation method are designed to integrate the advantages of CT’s density information and MRI’s soft tissue resolution. Finally, experiments conducted on the LiTS dataset demonstrate that the proposed model effectively leverages multimodal fusion to improve segmentation accuracy, achieving a dice score of 94% for multimodal inputs, thereby meeting clinical real-time requirements. These findings have significant implications for advancing the clinical application of precise liver tumor diagnosis and treatment.