Coronary artery disease detection using deep integration of vision transformer and extended LSTM from cardiac multi-modal images

Evaluating the presence and severity of coronary artery disease (CAD) in individuals is essential for establishing the most effective treatment approach. Computed tomography (CT) provides superior spatial resolution images of the heart and coronary arteries in a short time frame. Conversely, several issues arise in the examination of cardiac CT scans for markers of CAD. Research initiatives employ machine learning to provide elevated accuracy and reliable performance, hence overcoming current limitations of CT images such as their expensive, demandable sources and subjectivity. It enables enhanced imaging of the coronary arteries with increased spatial resolution. Contemporary approaches sometimes involve considerable computational costs and a wide range of parameters, especially in the absence of training data. Nonetheless, these models need considerable processing resources, a large amount of labeled data, and more memory capacity and have restricted effectiveness for sequential data. To overcome these challenges, we adopted a stringent and automated diagnostic method for the detection of CAD using the fusion of a vision transformer (ViT) and extended LSTM (xLSTM) named VXL. This adopted framework model utilizes patch-based feature extraction, which is effective for sequential data due to xLSTM, and results in reduced computing expenses. This study used 6624 CT angiography and 63,151 MRI images, which were pre-processed using morphological techniques, edge enhancement, Multidimensional-CLAHE (M-CLAHE), and Fourier transforms for the detection of CAD to produce contrast-enhanced and enhanced visualization. The experimentation is carried out for different combinations of preprocessing techniques with VXL to confirm the robustness and efficiency for CAD detection. The Morpho+M-CLAHE+VXL model achieved accuracies of 99.19%, 99.11%, and 99.23% for accuracy, recall, and precision on the MRI image dataset and 96.75%, 96.63%, and 96.53% for accuracy, recall, and precision on the CT angiography images, respectively. This study facilitates the successful identification of CAD using multi-modal image datasets and assists medical practitioners in administering appropriate therapy. Finally, we applied local interpretable model-agnostic explanations (LIME) and segmentation maps to ascertain the primary characteristics utilized in decision-making. (The scratch code is available at https://github.com/SAHEBGOUD/Coronary-Artery-Disease-sample-datasets-and-scratch-code )