Improved Two-Stage Transfer Learning Approach for ViT-Based Myocardial Infarction Detection

Myocardial infarction (MI) is a critical cardiovascular condition requiring precise diagnosis. Accurate segmentation of myocardial pathologies in cardiac magnetic resonance (CMR) images is essential for assessing myocardial viability. Transfer learning (TL) plays a key role in enhancing the performance of medical image analysis by leveraging knowledge from related tasks or domains. However, existing TL-based cardiac image segmentation methods frequently focus on transferring knowledge from segmentation tasks to address the challenge of limited labeled data. On the other hand, despite their potential, vision transformers (ViTs) remain underutilized in conjunction with TL techniques for cardiac image segmentation. This study introduces a two-stage TL approach called 2-TLViT for segmenting MI in CMR images, utilizing improved instance-based and network-based TL techniques with ViTs. Our method leverages knowledge from both classification and segmentation tasks to enhance segmentation performance. 2-TLViT includes three phases: preprocessing to extract the region of interest, pretraining the ViT model on cardiac classification and segmentation datasets, and fine-tuning on an MI diagnosis dataset. Extensive ablation and analysis experiments are conducted, to compare 2-TLViT against several state-of-the-art methods. The results demonstrate significant improvements in Dice similarity coefficient (Dsc) scores across various myocardial regions, including the myocardium (Myo), left ventricle (LV), right ventricle (RV), myocardial edema and myocardial scars. Notably, the proposed approach achieved the highest Dsc scores for these regions, significantly outperforming the baseline and other models. These results underscore the effectiveness of 2-TLViT, highlighting its potential for clinical application in automated myocardial pathology segmentation.