Adaptive Attention-Enhanced Transformer With Modified Graph Cuts (AAET-MGC) Algorithm for Cardiovascular Disease Diagnosis

One of the core prerequisites in the diagnosis and treatment of cardiovascular diseases is accurate segmentation of the myocardium walls and small vessels in cardiac MRI datasets. Traditional segmentation methodologies often do not tackle this complexity of cardiac structures and their variability across MRI images. This may further enhance overfitting and local minima problems during training of deep learning models for cardiac image segmentation. Against this backdrop, a new hybrid approach is introduced called Adaptive Attention Enhanced Transformer with Modified Graph Cuts. This methodology is devised to merge the local feature extraction using a Convolutional Neural Network with the global context awareness provided by the transformer encoders further enhanced by attention mechanisms; user interest is oriented toward the most relevant features. The AAET-MGC model integrates an adaptive energy function within a modified graph cuts framework with spatial and intensity information, aiming at refining the segmentation results. This hybrid approach uses the power of deep learning in an initial segmentation process and classical graph cuts for accurate and smooth boundary detection. The method is evaluated against a comprehensive cardiac MRI dataset with an accuracy of 97%, avoiding overfitting and local minima. The AAET-MGC approach outperforms competing methods in the accurate segmentation of thin and complex structures, conductive to better diagnostic accuracy for the related clinical applications.