This work explores if combining information from multiple Magnetic Resonance Imaging (MRI) modalities provides improved interpretation of brain biological architecture as each MR modality can reveal different characteristics of underlying anatomical structures. Structural MRI provides a means for high-resolution quantitative study of brain morphometry. Diffusion-weighted MR imaging (DWI) allows for low-resolution modeling of diffusivity properties of water molecules.
Structural and diffusion-weighted MRI modalities are commonly used for monitoring the biological architecture of the brain in normal development or neurodegenerative disease processes. Structural MRI provides an overall map of brain tissue organization that is useful for identifying distinct anatomical boundaries that define gross organization of the brain. DWI models provide a reflection of the micro-structure of white matter (WM), thereby providing insightful information for measuring localized tissue properties or for generating maps of brain connectivity. Multispectral information from different structural MR modalities can lead to better delineation of anatomical boundaries, but careful considerations should be taken to deal with increased partial volume effects (PVE) when input modalities are provided in different spatial resolutions. Interpretation of diffusion-weighted MRI is strongly limited by its relatively low spatial resolution. PVE's are an inherent consequence of the limited spatial resolution in low-resolution images like DWI.
This work develops novel methods to enhance tissue classification by addressing challenges of partial volume effects encountered from multi-modal data that are provided in different spatial resolutions. Additionally, this project addresses PVE in low-resolution DWI scans by introducing a novel super-resolution reconstruction approach that uses prior information from multi-modal structural MR images provided in higher spatial resolution.
The major contributions of this work include: 1) Enhancing multi-modal tissue classification by addressing increased PVE when multispectral information come from different spatial resolutions. A novel method was introduced to find pure spatial samples that are not affected by partial volume composition. Once detecting pure samples, we can safely integrate multi-modal information in training/initialization of the classifier for an enhanced segmentation quality. Our method operates in physical spatial domain and is not limited by the constraints of voxel lattice spaces of different input modalities. 2) Enhancing the spatial resolution of DWI scans by introducing a novel method for super-resolution reconstruction of diffusion-weighted imaging data using high biological-resolution information provided by structural MRI data such that the voxel values at tissue boundaries of the reconstructed DWI image will be in agreement with the actual anatomical definitions of morphological data.
We used 2D phantom data and 3D simulated multi-modal MR scans for quantitative evaluation of introduced tissue classification approach. The phantom study result demonstrates that the segmentation error rate is reduced when training samples were selected only from the pure samples. Quantitative results using Dice index from 3D simulated MR scans proves that the multi-modal segmentation quality with low-resolution second modality can approach the accuracy of high-resolution multi-modal segmentation when pure samples are incorporated in the training of classifier. We used high-resolution DWI from Human Connectome Project (HCP) as a gold standard for super-resolution reconstruction evaluation to measure the effectiveness of our method to recover high-resolution extrapolations from low-resolution DWI data using three evaluation approaches consisting of brain tractography, rotationally invariant scalars and tensor properties. Our validation demonstrates a significant improvement in the performance of developed approach in providing accurate assessment of brain connectivity and recovering the high-resolution rotationally invariant scalars (RIS) and tensor property measurements when our approach was compared with two common methods in the literature.
The novel methods of this work provide important improvements in tools that assist with improving interpretation of brain biological architecture. We demonstrate an increased sensitivity for volumetric and diffusion measures commonly used in clinical trials to advance our understanding of both normal development and disease induced degeneration. The improved sensitivity may lead to a substantial decrease in the necessary sample size required to demonstrate statistical significance and thereby may reduce the cost of future studies or may allow more clinical and observational trials to be performed in parallel.