3661 - Super-Resolving Any MRI Modality with a Patient-Specific Framework (SR-Any) to Accelerate Multi-Sequence MR-Guided Radiotherapy

Purpose/Objective(s):

In MR-guided radiotherapy, multiple MR sequences are usually desired for collective tumor/organ-at-risk contouring but challenging to implement due to time constraints. We developed a patient-specific super-resolution strategy (SR-Any) based on Implicit Neural Representation (INR). SR-Any leverages a patient-specific, high-resolution simulation or onboard MRI as prior information, to super-resolve any other sequences acquired via fast, low-resolution scans to save imaging time. Based on fast test-time learning (~1 min), SR-Any can super-resolve any MRI sequence without prior training, allowing the maximum flexibility while reducing hallucinations from population-based models.

Materials/Methods:

SR-Any constructs a patient-specific INR for super-resolution. The INR is a coordinator-based network, of which the inputs are imaging coordinates and outputs are the corresponding MR values at the input coordinates. INR essentially maps an image to a network-based implicit function, allowing flexible super-resolution through up-sampling input image coordinates. For SR-Any, to effectively incorporate patient-specific prior information, a local patch, extracted as a neighborhood of each queried image coordinate from a prior high-resolution MRI, was input along with the image coordinate to infer the corresponding MR value of the target image. The test-time training of SR-Any is supervised by the acquired low-resolution version of each target image, via a mean-squared-error loss between SR-Any-inferred MR values and acquired ones at the low-resolution image grids. The 'one-shot' SR-Any can be quickly trained onboard for prior and target MRIs of arbitrary sequences.

An in-house dataset of 3,322 MRIs covering 20 sequences from four anatomical regions (head-and-neck, chest, abdomen, and pelvis) was used to evaluate SR-Any, via metrics including the Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Learned Perceptual Image Patch Similarity (LPIPS). Additionally, the Segment Anything Model was employed to segment tumors in the super-resolved images, with DICE used to assess the segmentation performance.

Results:

SR-Any was evaluated on 162 distinct super-resolution tasks and compared with seven other methods. SR-Any achieved best performance across all metrics, with an SSIM of 0.921 ± 0.037, PSNR of 32.57 ± 4.41, LPIPS of 0.059 ± 0.028, and DICE of 0.688 ± 0.117. In contrast, the second-best method, RDN, achieved an SSIM of 0.917 ± 0.033, PSNR of 31.23 ± 2.54, LPIPS of 0.071 ± 0.030, and DICE of 0.649 ± 0.116.

Conclusion:

SR-Any reliably achieves super-resolution using a high-resolution prior MRI without the need for large-scale pretraining. This flexible approach enables super-resolution across any MR modality, enhances tumor segmentation, and accelerates the workflow of MR-guided radiotherapy.