143 - Computationally Efficient Low Field MRI Enhancement via Foundation Model Feature Engineering

Purpose/Objective(s): Low-field MRI is gaining popularity in imaging-guided radiotherapy due to its lower system cost and improved geometric fidelity, yet its clinical utility is limited by reduced signal-to-noise ratio and low image resolution. This work aims to enhance low-field MRI by developing a computationally efficient deep learning framework that leverages pretrained foundation model features to predict high-quality images.

Materials/Methods: T1- and T2-weighted brain MRI acquired on a low-field (0.3 T) scanner were investigated. Low-resolution inputs were generated through 3× k-space subsampling. High-quality ground truth was obtained by repeating full k-space acquisition three times, followed by signal averaging to suppress noise. A pretrained vision transformer from an image segmentation foundation model was used to extract multi-level global representations from the inputs. These features were adapted through a lightweight adaptor network that aligns transformer embeddings with convolutional feature maps in both spatial resolution and channel dimensions. The adapted global features were fused with locally extracted convolutional features and decoded via a convolutional decoder with skip connections to generate enhanced outputs. The model was trained using T1-weighted MRI from 128 subjects and subsequently fine-tuned using T2-weighted MRI from 32 subjects to evaluate its transfer learning capability. Model performance was evaluated for both denoising and 3× super-resolution (SR) using 32 independent testing subjects.

Results: The proposed method enhanced image quality, leading to peak signal-to-noise ratio (PSNR) improvements of 5.66 and 5.11 dB for 3× subsampled T1 and T2 low-field MRI, respectively, which were significantly (p<0.001) superior to both CNN- and transformer-based state-of-the-art (SOTA) methods. Compared with the best-performing SOTA methods (RESTORMER and SwinIR), the proposed method reduced GPU memory usage by 64.6% and inference time by 58.1%.

Conclusion: By engineering pretrained foundation model features, high-quality low-field MRI can be generated from noisy and low-resolution inputs without large-scale network retraining. This computationally efficient approach enables practical deployment of AI-enhanced low-field MRI to support target and organ-at-risk visualization in MRI-guided radiotherapy.

Abstract 143 - Table 1. Quantitative comparison between inputs, model outputs and the ground truth