Self-Supervised Learning with Unsupervised Motion Correction for Fitting IVIM Model
Feilin Deng, Shangxuan Li, Baoer Liu, Yikai Xu, Zhou Ping Wu · 2024
The parameter maps of Intravoxel incoherent motion (IVIM) DWI models play an important role in clinical diagnosis and treatment, and current fitting methods based on deep learning have achieved promising results, including fully-supervised learning and self-supervised learning methods. However, due to the presence of motion in the IVIM sequence caused by heartbeat or respiration, current methods have not considered the impact of motion in the IVIM sequence on fitting performance. In this work, we propose embedding motion correction into self-supervised learning for IVIM parameter fitting of IVIM sequences, by designing an end-to-end deep network that simultaneously performs both motion correction and IVIM parameter fitting tasks. Our idea comes from the observation that motion correction and IVIM parameter fitting are two interrelated and mutually reinforcing tasks. Importantly, the motion correction between sequences is unsupervised learning, similar to parameter fitting in self-supervised learning, which does not require any training data, greatly facilitating network design and clinical application. Experiments of clinical data have shown that the proposed motion correction and parameter fitting network of multi-task learning outperform the single task parameter fitting network in terms of parameter fitting performance, and also outperform the parameter fitting performance while two independent tasks are run separately.