High-fidelity Ultra-fast Diffusion Tensor Imaging in Stroke Patients Using Transfer Learning

Motivation: Deep learning shows potential for accelerating diffusion tensor imaging (DTI) without compromising image quality, but its feasibility has not been comprehensively evaluated for clinical applications. Goal(s): Achieve fast, high-quality DTI for clinical use with minimal training data. Approach: We adopt a U-Net to map low-SNR 6-direction diffusion data to high-SNR data with more diffusion directions. The U-Net is pretrained on UKB data and fine-tuned with limited clinical data from stroke patients. Results: Evaluations on clinical data reveal our model effectively reduces scan-time from 9.7 minutes to 1 minute while producing high-SNR diffusion images, accurate diffusion metrics, and high-quality tractography. Impact: Our work supports and promotes the feasibility of deep learning approach to benefit the clinical adoption of DTI for diagnosis and/or pre-surgical planning in scenarios where the scan time is extremely limited (e.g., for stroke patients).