DORIS: A diffusion MRI-based 10 tissue class deep learning segmentation algorithm tailored to improve...

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning in Medical Imaging

Background:

• Tractography algorithms rely on accurate tissue segmentation (white matter, gray matter, CSF) for reliable results.
• Current methods often use T1-weighted images, requiring challenging registration to diffusion space, leading to inaccuracies.
• Diffusion-based segmentation is needed for improved anatomical priors in tractography.

Purpose of the Study:

• To introduce DORIS, a deep learning algorithm for tissue segmentation directly in native diffusion-weighted image (DWI) space.
• To evaluate DORIS's performance against existing methods and assess its impact on diffusion MRI tractography.
• To provide a robust and accurate segmentation tool for advanced neuroimaging analyses.

Main Methods:

• Developed DORIS, a deep learning model trained on 1,000 subjects (ages 22-90) using diverse DWI data from public databases.
• Employed a silver standard strategy using Freesurfer outputs registered to DWI space for training and validation.
• Quantitatively compared DORIS segmentation maps with Freesurfer and FSL-fast, and evaluated tractography outcomes.

Main Results:

• DORIS accurately segments 10 tissue classes, including subcortical structures, directly from DWI.
• Segmentation maps from DORIS showed high accuracy compared to established methods.
• Tractograms generated using DORIS priors exhibited longer mean streamline length and reduced anatomical implausibility.

Conclusions:

• DORIS offers a fast, accurate, and reproducible solution for DWI-based tissue segmentation.
• Utilizing DORIS segmentation improves the quality and anatomical validity of diffusion MRI tractography.
• This approach overcomes limitations of T1-based segmentation and FA-based tractography, advancing neuroimaging analysis.