Transformer-based RAD-FE-MCDFN Framework for Cardiac Adipose Tissue Segmentation and Disease Diagnosis

Aligning of multiarmed bandit strategies with GAN architectures in prior works exhibited encouraging segmentation efficacy; however, it also introduced considerable computational complexity and necessitated extensive hyperparameter optimization elements that may impede implementation in resource-limited clinical environments. To tackle these issues and enhance cardiac fat analysis, we offer an innovative hybrid ensemble-transformer framework that combines the Residual Attention-based Deep Feature Extraction (RA-FE) method with the Multi-Channel Data Fusion Network (MCDFN).RAD-FE employs multi-scale convolutional layers enhanced by attention mechanisms and residual learning to extract unique, fine-to-coarse characteristics from cardiac imaging data, whereas MCDFN integrates these imaging features with clinical information using a channel-wise attention method. Our methodology surpasses current state-of-the-art techniques, as evidenced by assessments on a diverse dataset, which yield an F1-score of 95.0%, sensitivity of 93.8%, specificity of 96.5%, and accuracy of 95.4%. This varied strategy aids to early detection and better control of cardiovascular diseases, boosting the accuracy and reliability of cardiac adipose tissue identification.