YOLO-M: Integrating Spatial Attention and Multi-Scale Learning for Robust Liver Tumor Detection

An early diagnosis of hepatocellular carcinoma is among the most lethal cancers of the liver, via CT imaging is vital to the prompt intervention. Nevertheless, localization of lesions is impeded as a result of hepatic complicated anatomy, low contrast and variegated morphology of lesions. We should overcome these challenges in order to achieve the following present YOLO-M, a liver-focused object detection system which enhances the YOLO design with a backbone using a two-path, a Multi-Scale Liver Feature Attention (LFA) module and a Liver Region Attention Module (LRAM). YOLO-M also has a disease pattern detection head and an anchor that is personalized box design adapted to the hepatic lesions. YOLO-M is evaluated on the HCC-TACE-Seg dataset and achieves an output mAP@0.5 of 0.975, Dice score of 0.9056, accuracy of 98.2%, precision of 97.44%, recall of 84.59% and F1-score of 90.56% which surpasses state-of-the-art classifiers such as YOLOv5, HFCNN and TransUNet. These findings indicate that YOLO-M is a powerful and effective tool which can be used to detect real-time liver lesion detection and segmentation.