Deep Learning-Based Quantitative T1 Mapping of the Brain for a Digital MRI Phantom: preliminary study

The advancement of neural networks for MRI data processing encounters a significant challenge: the acquisition of large-scale, real-world datasets. A promising solution to this limitation is the development of realistic digital MRI phantoms. In this study, we focused on generating quantitative T1 maps of the human brain for a digital MRI phantom using deep learning approach, specifically denoising diffusion probabilistic models (DDPM). To train our model, we constructed a dataset of quantitative T1 maps derived from anatomical brain masks and relaxation times reported in the literature. Using a similar methodology, additional quantitative maps—including T2, T2*, and proton density—were generated and incorporated into a comprehensive digital phantom. T1-weighted MRI images were subsequently synthesized using these phantoms in combination with a real turbo-spin echo pulse sequence and an MRI simulator. A diffusion model was trained to predict T1 maps based on T1-weighted MR images. The quality of the T1 maps predicted by the trained DDPM was assessed by comparing them with the original maps using the structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) yielding values of 0.71 and 19.22 respectively. The generated T1 maps were used to create new T1-weighted images, which were compared with their original counterparts, yielding SSIM and PSNR values of 0.78 and 18.48, respectively. These results suggest that this approach can be extended to generate other quantitative maps. This, in turn, enables various applications, including pulse sequence optimization, evaluation of reconstruction algorithms in the absence of real quantitative MRI data.