Enhancing medical image super-resolution through generative adversarial networks and transformer-based joint feature learning
R. Balamanigandan, R Mahaveerakannan, R. Yuvarani, Pankaj Dadheech, Sanwta Ram Dogiwal, Chour Singh Rajpoot, Digvijay Pandey · 2024
Super-resolution reconstruction in medical imaging has become increasingly crucial for obtaining high-quality images while minimizing radiation exposure, particularly in low-field magnetic resonance imaging (MRI). However, achieving accurate super-resolution remains challenging due to the intricate nature of medical images and their high diagnostic requirements. In this paper, we propose a deep learning-based strategy for reconstructing medical images from low resolutions, employing a novel integration of Transformer and Generative Adversarial Networks (T-GANs). In comparison to established metrics such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM), our T-GAN approach consistently achieves optimal performance and demonstrates superior texture feature recovery in the super-resolution reconstruction of MRI scans, specifically focusing on knee and abdominal images. Our methodology integrates Transformer and GANs to create a robust system capable of reconstructing medical images from low-resolution inputs. Furthermore, we introduce a novel multi-task loss function, wherein we assign weights to content loss, adversarial loss, and adversarial feature loss. This loss function optimizes the training of our T-GAN model, allowing it to produce superior super-resolution results. Our proposed T-GAN approach demonstrates remarkable performance improvements compared to conventional metrics such as PSNR and SSIM.