Purpose: Deep learning (DL)-based MRI super-resolution (SR) reconstruction has recently been receiving attention due to the significant improvement in spatial resolution compared to conventional SR techniques. However, there remain several barriers to the effective implementation of these approaches in a clinical setting. Low-resolution (LR) MRIs captured in the clinic contain complex tissues obfuscated by noise that are difficult for a simple DL framework to reconstruct. Moreover, training a robust SR network requires abundant, perfectly matched pairs of LR and high-resolution (HR) images that are often clinically unavailable. We propose here a patient-specific SR approach utilizing the Intentional Deep Overfit Learning (IDOL) framework that demonstrates a significant improvement in SR MRI quality using only a few training data sets.
Methods: IDOL-based SR is an intentional application of overfitting that leverages patient-specific prior information. Training in the IDOL framework involves two stages:1)conventional training of a generalized SR model with a training set of LR and HR MRIs of 4 human volunteers, and 2)using the inhale respiratory phase of a 5th human subject as prior information, the generalized model is trained further to achieve the IDOL model that is applied to the exhale respiratory phase of the 5th subject. For testing, step2 was repeated for subjects 5-8 using the same generalized model as a starting point.
Results: The proposed method was evaluated on a cohort of 4 volunteers. The generalized model (step1 training only) achieved a normalized mean absolute error (MAE) of 446.5 and peak signal-to-noise ratio (PSNR) of 31.33. By adopting the personalized IDOL model in each test case, the MAE and PSNR improved to 321.5 and 34.72, respectively.
Conclusion: In this study, we have successfully applied the IDOL framework to the MRI super-resolution task. Our approach is widely applicable to common tasks in radiotherapy and can be utilized for MRI-guided ART.