Diffusion model-based augmentation using asymmetric attention mechanisms for cardiac mri images

Cardiovascular diseases constitute one of the leading causes of mortality worldwide, claiming approximately 19.8 million lives annually. Although cardiac magnetic resonance (CMR) imaging is considered the gold standard for diagnosing heart diseases, the development of deep learning-based diagnostic systems faces significant challenges due to expert labeling costs, ethical constraints, and inadequate representation of rare pathologies in creating large-scale labeled datasets. This thesis proposes a novel synthetic image generation method based on a diffusion model with asymmetric attention mechanisms to address the data scarcity problem in cardiac MR imaging. The proposed approach is built upon the Denoising Diffusion Probabilistic Model (DDPM) framework and incorporates multi-head self-attention mechanisms strategically placed in the last two layers of the encoder and the first two layers of the decoder within the UNet architecture. This asymmetric attention placement achieves an optimal balance between computational efficiency and anatomical structure preservation. The model was trained on the OCMR (Open Cardiac MR) dataset obtained from 53 healthy volunteers, working with bSSFP sequence images at 128×128 pixel resolution. Comprehensive experimental evaluations demonstrated that the proposed model exhibits superior performance compared to existing methods. The model achieved an FID score of 77.78, significantly outperforming VAE (325.26), WGAN-GP (235.51), and StyleGAN2-ADA (117.70). Structural similarity metrics of 0.720 SSIM and 0.925 MS-SSIM confirm the high anatomical consistency of the generated images. Ablation studies showed that the asymmetric attention placement provides an 8.8% improvement in FID score compared to uniform distribution. A clinical Turing test conducted by three experienced radiologists yielded a 60% accuracy rate, demonstrating the difficulty in distinguishing synthetic images from real ones. Furthermore, no statistically significant differences were found between real and synthetic images in terms of pixel ratios and shape features of the left ventricle, right ventricle, and myocardium structures (p > 0.05). This thesis presents an original contribution to asymmetric attention mechanism optimization for diffusion models in cardiac MR image synthesis and demonstrates the clinical applicability of the generated synthetic data through a comprehensive evaluation framework. The results indicate that the proposed approach holds significant potential in areas such as data augmentation, privacy-preserving data sharing, and medical education.