Author: Jie Deng, Yunxiang Li, You Zhang 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Laboratory, Department of Radiation Oncology, UT Southwestern Medical Center 🌍
Purpose: Magnetic Resonance Imaging (MRI) has exceptional soft tissue contrast and an essential role in radiotherapy. The introduction of clinical MR-LINACs has enabled adaptive radiotherapy (ART) using onboard MRI for each treatment, maximizing treatment accuracy. However, due to hardware limitations, patient comfort, and time constraints associated with ART, the spatial resolution of onboard MRI is often compromised, particularly when acquiring multiple MRIs with different contrast weightings.
Methods: Our model is based on Implicit Neural Representation (INR), which learns a neural network-based mapping from image coordinates to signal values. Since coordinate values are continuous, the model can be trained using sparse low-resolution coordinates while performing inference on full high-resolution coordinates to achieve super-resolution. We propose a Universal Anatomical Mapping and Patient-Specific Prior INR (USINR) framework, which integrates database-level general learning with patient-specific fine-tuning. The general learning phase leverages an existing high-resolution MRI database to instill population-level knowledge, providing a fast INR initialization based on a patient-specific prior high-resolution MRI. Patient-specific training then fine-tunes the initialized INR based on each patient’s low-resolution MRIs during test time to minimize the risk of hallucinations, all within <30 seconds.
Results: USINR was evaluated on three super-resolution tasks using the IXI and BRATS datasets, achieving the best performance in all tasks. For example, in a super-resolution task on IXI, USINR achieved average±s.d. SSIM, PSNR, and LPIPS scores of 0.964±0.004, 35.064±0.547, and 0.012±0.002, respectively, compared to MCINR's (state-of-the-art) scores of 0.943±0.006, 31.466±0.948, and 0.049±0.008. Our method is found robust to substantial anatomical/intensity changes between patient-specific prior high-resolution MRI and onboard low-resolution MRI, thanks to the patient-specific fine-tuning.
Conclusion: By combining general anatomical mapping knowledge with patient-specific INR fine-tuning, USINR offers a novel and reliable approach to MRI super-resolution. With minimal processing time, it effectively balances the need for image quality with the efficiency requirements of MRI-guided ART.