Incorporating spatial information in deep learning parameter estimation with application to the intravoxel incoherent motion model in diffusion-weighted MRI

In medical image analysis, the utilization of biophysical models for signal analysis offers valuable insights into the underlying tissue types and microstructural processes. In diffusion-weighted magnetic resonance imaging (DWI), a major challenge lies in accurately estimating model parameters from the acquired data due to the inherently low signal-to-noise ratio (SNR) of the signal measurements and the complexity of solving the ill-posed inverse problem. Conventional model fitting approaches treat individual voxels as independent. However, the tissue microenvironment is typically homogeneous in a local environment, where neighboring voxels may contain correlated information. To harness the potential benefits of exploiting correlations among signals in adjacent voxels, this study introduces a novel approach to deep learning parameter estimation that effectively incorporates relevant spatial information. This is achieved by training neural networks on patches of synthetic data encompassing plausible combinations of direct correlations between neighboring voxels. We evaluated the approach on the intravoxel incoherent motion (IVIM) model in DWI. We explored the potential of several deep learning architectures to incorporate spatial information using self-supervised and supervised learning. We assessed performance quantitatively using novel fractal-noise-based synthetic data, which provide ground truths possessing spatial correlations. Additionally, we present results of the approach applied to in vivo DWI data consisting of twelve repetitions from a healthy volunteer. We demonstrate that supervised training on larger patch sizes using attention models leads to substantial performance improvements over both conventional voxelwise model fitting and convolution-based approaches.