GAN Synergy: Transfer Learning-based Annotated Left Atrium MR Imaging Generation with Paired and Unpaired Models

This study presents a novel approach for generating left atrium Magnetic Resonance Imaging (MRI) by combining Paired and Unpaired Generative Adversarial Network (GAN) models. The aim is to address the limited dataset available for diagnosing and treating cardiovascular illness. Early detection and treatment of cardiovascular illnesses, which are the leading cause of death globally, need the use of advanced imaging techniques. The need for annotated medical imaging datasets in cardiology poses challenges for constructing deep learning models due to high collection costs, privacy concerns, and the limited availability of specific data. The study employs transfer learning from a pre-trained U-Net model to enhance Pix2Pix (paired GAN) and transfers it to CycleGAN (unpaired GAN) models to overcome dataset constraints. Our technique leverages the high fidelity of Pix2Pix and the adaptability of CycleGAN to create more annotations, even without extensive annotated datasets. The proposed methodology involves transfer learning between GAN models, enhancing the quality of generated annotations based on the evaluation method with Structural Similarity Index Measure (SSIM) and Peak Signal to Noise Ratio (PSNR), and improving the stability of the training process. The proposed method shows that the Pix2Pix model exhibits an average SSIM of 0.92 and PSNR of 29.03 compared to the actual and generated outputs. The CycleGAN model exhibits an average SSIM of 0.84 and a PSNR of 25.74 compared to the real input and generated output.