Integrated Cardiac Diagnosis and Prognosis using an Optimized Asynchronous Convolution–Propagation Transformer Network with Quaternion Feature Learning

Abstract

Heart diseases continue to pose the most significant threat to people's lives globally; hence, effective and reliable cardiovascular magnetic resonance imaging analysis is key to early diagnosis and predicting patient outcomes. Deep learning techniques have been instrumental in developing models that can automatically analyze cardiovascular MRIs by recognizing complex anatomical and physiological features. Nevertheless, current systems face limitations in sensitivity to noise, insufficient modeling of image features, and ineffective incorporation of both local and global interactions in multiple slice heart MRIs. In this study, a novel system of integrated cardiac diagnosis and prognosis using optimized asynchronous convolution–propagation transformer network with quaternion feature learning (ACPTN-SBFO) was developed through the process of data collection (CAD-CMR dataset), preprocessing with structure-aware adaptive bilateral texture filter (SABTF) to enhance edges and suppress noise, extraction of features with short time quaternion quadratic phase Fourier transform (STQQPFT) to incorporate spatial–frequency and phase, classification with asynchronous convolution–propagation transformer network by merging convolution-transformer (CT) with asynchronous propagation attention network (APAN), and optimization with sharpbelly fish optimization (SFO). Overall, high accuracies of 99.21% were obtained with 99.08% precision, 99.15% recall, and 99.11% F1-score.