High-Resolution 7T-like MRI Images Generation from 3T MRI Scans Using Multi-Scale Residual CycleGAN
Franklin Burhagohain, Shovan Barma · 2024
This work aims to generate high-resolution 7T-like MRI from 3T MRI scans utilizing Multi-Scale Residual CycleGAN (MSR-CycleGAN). Image generation tasks within the Generative Adversarial Network (GAN) framework have yielded notable outcomes. However, existing studies predominantly focus on the Super-Resolution GAN, which demands paired datasets, limiting its applicability, as collecting paired dataset is very challenging, particularly in medical domain. It can be overcome by using MSR-CycleGAN, which is designed to work in unpaired data settings. To ensure suitable MRI scans, 100 slices from each of three distinct anatomical planes-axial$(A_{x})$, coronal$(C_{r})$, and sagittal$(S_{g})$-across 10 subjects were selected. Which were subsequently processed through MSR-CycleGAN using two distinct approaches: (a) Merged Planes, where planes were combined$\left(A_x C_r S_g\right)$; and (b) Separate Planes, where each plane was processed individually as$A_{x}, C_{r}$, and$S_{g}$. For validation, benchmark UNC 3T-7T dataset has been taken into account. The evaluation was conducted by measuring the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). In a comparative evaluation, state-of-the-art methods, including SRGAN, have been considered to assess performance. The findings reveal that MSR-CycleGAN surpasses SRGAN, with the$S_{g}$input achieving PSNR and SSIM values of 27.56 and 0.82, respectively. This analysis also indicates that focusing on individual planes may be more effective for MRI generation when employing GAN-based techniques.