Author: Kofi M. Deh, Tamas Jozsa, Tsang-Wei Tu 👨🔬
Affiliation: Cranfield University, Howard University Hospital, Howard University 🌍
Purpose: To enhance the quality of hyperpolarized (HP) 13C magnetic resonance images by integrating deep learning with perfusion modeling.
Methods: A convolutional neural network (CNN) and a superresolution generative adversarial network (SRGAN) were developed to interpolate HP 13C images to high-resolution anatomical images. The models were trained on 75 synthetic volumetric brain perfusion images, generated using a porous media model of the brain. Training data were augmented with Rician noise and geometric transformations. The performance of the trained networks was evaluated by interpolating low-resolution synthetic test images and comparing the interpolated outputs to ground truth data with metrics such as peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). The neural networks were also applied to clinical HP 13C images, and the results were analyzed by comparing regions of interest (ROIs) and volumetric renderings.
Results: The neural networks outperformed traditional interpolation methods on the test synthetic perfusion dataset, achieving higher PSNR and SSIM values. The neural network interpolated HP 13C images also displayed reduced noise and improved structural clarity.
Conclusion: Deep learning models trained on synthetic perfusion data significantly enhance the quality of interpolated HP 13C images compared to existing techniques.