Abstract
The classification of the current density vector map (CDVM) reconstructed from magnetocardiogram (MCG) is an important indicator for assessing cardiac function and state in clinical diagnosis. Given the limited widespread application of MCG, research on CDVM often encounters challenges such as scarcity of data and difficulties in judgment. There is growing interest in computer-aided methods to assist physicians in analyzing cardiac cases using CDVM. This paper proposes a deep learning-based approachto classify the CDVM. To address the issue of insufficient processed MCG data, data augmentation is carried out by adding noise, making predictions based on auto regressive integrated moving average (ARIMA) model, and utilizing interpolation methods. A transformer hybrid residual network is then employed to classify the CDVM across categories 0 to 4, with transfer learning incorporated into the network structure to initialize model parameters, and the self-attention mechanism of the Transformer enhancing the feature extraction capability. This method achieved a classification accuracy of 97.52%, outperforming previous deep learning approaches, exhibiting both high precision and efficiency. Furthermore, its high scalability ensures that it will continue to meet the evolving needs of physicians as CDVM datasets undergo continuous expansion.