Automatic identification of ischemia using lightweight attention network in PET cardiac perfusion imaging

Abstract

Ischemic disease, caused by inadequate blood supply to organs or tissues, poses a significant global health challenge. Early detection of ischemia is crucial for timely intervention and improved patient outcomes. Myocardial perfusion imaging with positron-emission tomography (PET) is a non-invasive technique used to identify ischemia. However, accurately interpreting PET images can be challenging, necessitating the development of reliable classification methods. In this study, we propose a novel approach using MS-DenseNet, a lightweight attention network, for the detection and classification of ischemia from myocardial polar maps. Our model incorporates the squeeze and excitation modules to emphasize relevant feature channels and suppress unnecessary ones. By effectively utilizing channel interdependencies, we achieve optimum reuse of interchannel interactions, enhancing the model's performance. To evaluate the efficacy and accuracy of our proposed model, we compare it with transfer learning models commonly used in medical image analysis. We conducted experiments using a dataset of 138 polar maps (JPEG) obtained from 15O_H2O stress perfusion studies, comprising patients with ischemic and non-ischemic condition. Our results demonstrate that MS-DenseNet outperforms the transfer learning models, highlighting its potential for accurate ischemia detection and classification. This research contributes to the field of ischemia diagnosis by introducing a lightweight attention network that effectively captures the relevant features from myocardial polar maps. The integration of the squeeze and excitation modules further enhances the model's discriminative capabilities. The proposed MS-DenseNet offers a promising solution for accurate and efficient ischemia detection, potentially improving the speed and accuracy of diagnosis and leading to better patient outcomes.