SAM-MyoNet: A fine-grained perception myocardial ultrasound segmentation network based on segment anything model with prior knowledge driven

The automatic segmentation of myocardial ultrasound images is critical for the early diagnosis of cardiac diseases and the assessment of cardiac function. However, this task remains highly challenging due to the poor image quality of cardiac ultrasound and the complex structural morphology. In recent years, Segment Anything Model (SAM) and its derivative algorithms have shown superior performance in various complex segmentation tasks due to their extensive prior knowledge and powerful inference ability. Despite its strength in perceiving and inferring the myocardial edges, the segmentation accuracy of SAM still needs to be improved. To better complete this task, we propose a novel network SAM-MyoNet based on SAM, which has three main contributions. First, we introduce the information augmentation and driving module (IADM), which utilizes prior knowledge to augment the original dataset and drive SAM for feature extraction, thereby generating high-quality preliminary segmentation predictions. Second, we introduce the fine-grained feature perception module (FFPM), which employs a dual-branch attentional mechanism to refine segmentation based on preliminary results. One branch enhances fine-grained features, while the other maintains overall morphological perception, further improving segmentation accuracy. Third, we incorporate shape supervision to improve the model’s learning of myocardial shape characteristics. To comprehensively evaluate the network’s performance, we conducted extensive experiments across four different myocardial ultrasound datasets. The results show that our SAM-MyoNet outperforms the current state-of-the-art (SOTA) methods. Furthermore, we validate its generalizability using multi-center data. Our code is released at https://github.com/yingyuhan/SAM-MyoNet.