Motivation: To eliminate the reliance on abundant high-quality reference maps and to address the inconsistent distribution induced from external training data in learning-based IVIM parameter mapping. Goal(s): To develop a scan-specific unsupervised learning method that accurately maps IVIM parameters from noisy diffusion-weighted images without depending on external training data. Approach: The proposed method employs a physics-informed mechanism and convolutional implicit neural representation to generate IVIM parameter maps solely from diffusion-weighted images themselves. Results: Numerical phantom and in vivo human brain results demonstrate that our proposed method generates accurate IVIM parameter maps with improved quality. Impact: Our scan-specific unsupervised method, employing physics-informed convolutional implicit neural representation, has successfully achieved accurate IVIM parameter mapping without relying on reference maps and external data during network training.
Wang et al. (Tue,) studied this question.