DW-MRI Microstructure Model of Models Captured Via Single-Shell Bottleneck Deep Learning

Abstract

In the last two decades, numerous methods have been proposed to capture biological meaning from observed diffusion-weighted magnetic resonance imaging (DW-MRI) signals each addressing recovery of specific tissue properties (e.g. pathology based, volume fraction of tissues). Generically, specific tissue properties are recovered via a category of methods termed as tissue compartment modeling. Many recent compartmental approaches require two or more diffusivity shells. We hypothesize existence of a low dimension common representation for a wide variety of commonly used microstructural measures in common use. To test this hypothesis, we constructed 13 voxel-wise measurements from 5 distinct model-based approaches and used a multi-task deep convolutional network with a variable width bottleneck to infer these metrics from empirical DW-MRI. This is the first study to use data-driven exploration to map a common basis among DW-MRI modeling approaches. We propose to capture a compact feature space in the form of a bottleneck that preserves common features to all methods and retrieve information from single shell DW-MRI. We train on 3D patches of 40 Human Connectome Project (HCP) subjects ($\sim$27 million patches) where input is based on single shell DW-MRI and ground truth is estimated from all three shells of HCP data. We validate on 24 subjects and test on 25 subjects. The error reported for 5 microstructure methods on test set: 4.0%, 2.7%, 6.3%, 4.5% and 3.6%. We find that 6 features in the bottleneck layer efficiently capture an intrinsic feature space for the range of DW-MRI metrics explored.