Pretraining Improves Deep Learning Based Tissue Microstructure Estimation

Diffusion magnetic resonance imaging (dMRI) is commonly used to noninvasively estimate brain tissue microstructure, which provides important biomarkers for studying the structural changes of the brain. Due to the constraint of imaging time, the quality of dMRI scans can be limited by the number of diffusion gradients and the spatial resolution, and deep learning based approaches have been developed to provide high-quality estimation of tissue microstructure from the low-quality diffusion signals. In existing deep learning based methods, the estimation models are trained from scratch. However, it has been shown in various tasks that pretraining can improve the performance of deep networks, and it may also be used to improve deep learning based tissue microstructure estimation because there are abundant publicly available high-quality dMRI datasets. Moreover, for many datasets where acquisitions of a large number of high-quality training dMRI scans are not convenient, pretraining may also allow deep learning based methods to be applied with only a small number of training samples. Motivated by these potential benefits of pretraining, in this work, we explore whether pretraining improves deep learning based tissue microstructure estimation and how to achieve such improvement. Suppose we are given an auxiliary dataset with high-quality dMRI scans for pretraining and a target dMRI dataset of interest with a certain amount of high-quality training data. To generate inputs for pretraining, the diffusion signals of the auxiliary dataset are first downsampled in the spatial domain. Then, since the acquisition scheme is usually different between the two datasets, we interpolate the downsampled signals in the q-space using a dictionary-based signal representation. Finally, the downsampled and interpolated diffusion signals are used for pretraining the estimation network and the pretrained model is fine-tuned with the training data of the target dataset. Experiments were performed on brain dMRI scans, where we show that pretraining leads to improved accuracy of tissue microstructure estimation under different settings and may reduce the burden of training data acquisition.