Improving reconstruction of patient-specific abnormalities in AI-driven fast MRI with an individually adapted...

Artificial intelligence-based reconstruction of magnetic resonance imaging (MRI) from undersampled k-space data (namely, fast MRI) is not yet tailored for defining target and critical organs in radiotherapy (RT) planning. Previously proposed methods are often limited in reconstructing anatomic abnormalities, including tumors and surgical defects, because the population-based priors from training datasets do not necessarily provide such patient-specific information.

This study aims to advance deep-learning based MRI for RT planning by improving its ability to capture individualized abnormalities. This is achieved by incorporating previously acquired patient-specific MRI as an additional prior utilizing a denoising diffusion model.

In generative denoising diffusion models, an artificial neural network is typically trained to produce a random, realistic but fake image from a Gaussian noise. The generated images reflect the distribution of the population-based prior images which were used for training the network. When the denoising process is conditioned on undersampled k-space data, a fully sampled image, that is consistent with the patient of interest, is estimated. Whereas the conventional diffusion models use a fixed pre-trained neural network for all inference cases, a recent model adapted the network by the k-space data to improve its performance. We extended this approach to further adapt the network: it was first tuned to increase the similarity between the estimated image and the patient-specific prior image, followed by updating the network to maximize data consistency in the k-space. The proposed method was evaluated using fully sampled T1-weighted images from 73 pediatric and young adult patients who received RT for brain tumors, which were split into training (n = 58) and inference (n = 15) datasets. For each inference case, an additional image acquired at a median of 35 days from the inference image was used for the patient-specific prior.

With a simulated four-fold undersampling, the proposed model incorporating patient-specific prior resulted in a higher accuracy than the previous adaptive diffusion model without such prior (SSIM, 0.965 ± 0.010 vs. 0.938 ± 0.014; PSNR, 28.5 ± 1.5 vs. 25.8 ± 1.7; p < 0.001). Another diffusion model without any patient-specific adaptation showed the worst performance (SSIM, 0.912 ± 0.015; PSNR, 22.3 ± 1.1; p < 0.003). The superior accuracy of the proposed model was pronounced in the tumor and surgical defect. The PSNR within GTV was significantly higher than those of the other diffusion models (30.01 ± 1.9 vs. 27.43 ± 2.5 and 24.04 ± 2.7; p < 0.001).

Incorporating patient-specific prior in addition to population-based prior can significantly improve the accuracy of deep learning-based fast MRI, particularly in regions with anatomic abnormalities, thereby facilitating target definition for RT planning. The proposed model would be particularly useful for accelerating MRI simulation following diagnostic imaging or repeat MRI for MR-guided on-treatment adaptation.