Cross-granularity multi-task network for ischemia diagnosis and defect detection in the myocardial perfusion imaging

Myocardial perfusion imaging has been considered the reference standard for detection of myocardial ischemia in coronary heart disease. In clinical practice, doctors typically estimate the individual ischemic scores for 17 myocardial segments and then diagnose myocardial ischemia based on the scores for all segments considered together. However, prior works are generally limited to the diagnosis of myocardial ischemia and ignore the correlation between myocardial ischemia and the myocardial segment showing the defect. In this paper, we explore intra- and inter-granularity relationships between myocardial perfusion imaging tasks of different granularities and construct a multi-task learning framework to jointly learn these tasks. To this end, we propose a cross-granularity multi-task network, namely CGMT-Net. We present a task-specific module based on the attention mechanism to make the feature map task-specific. To explore the intra- and inter-granularity relationships among tasks, we propose a cross-granularity fusion module for the integration and transmission of task-specific information. We present a large dataset containing 1098 myocardial perfusion imaging sets paired with their diagnostic reports. Extensive experiments on the dataset demonstrate the superiority of our CGMT-Net over other methods. Furthermore, we performed ablation experiments that show the rationality and effectiveness of our .