DeepDTI: High-fidelity six-direction diffusion tensor imaging using deep learning.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• Diffusion Tensor Imaging (DTI) is crucial for mapping brain microstructure and connectivity.
• Conventional DTI requires long scan times due to extensive angular sampling, limiting clinical and research applications.
• High-quality DTI is essential for accurate tractography and neuroscientific studies.

Purpose of the Study:

• To introduce DeepDTI, a novel processing framework to minimize data requirements for DTI.
• To enable high-quality DTI acquisition with significantly reduced scan times.
• To improve the feasibility of DTI, tractography, and white-matter analysis in clinical and research settings.

Main Methods:

• Developed DeepDTI, a data-driven supervised deep learning framework using a 10-layer 3D convolutional neural network (CNN).
• Input data includes a non-diffusion-weighted image (b=0), six diffusion-weighted images (DWIs) along optimized directions, and T1/T2-weighted images.
• The CNN learns residuals between input and high-quality output images, enabling tensor fitting and DTI metric generation.

Main Results:

• DeepDTI achieves 3.3-4.6x acceleration, producing DTI metrics comparable to conventional DTI with significantly fewer images (2 b=0, 21-30 DWIs vs. 18 b=0, 90 DWIs).
• The framework demonstrates robust and rotationally-invariant estimation of DTI metrics, outperforming a state-of-the-art denoising algorithm.
• Accurate identification of 20 major white-matter tracts was achieved, with tract core distances of 1-1.5 mm compared to ground truth.

Conclusions:

• DeepDTI effectively leverages deep learning and diffusion MRI physics to enable rapid, high-quality DTI acquisition.
• The framework significantly reduces data requirements, making advanced neuroimaging analyses more accessible.
• DeepDTI holds promise for wider adoption in neuroscientific research and clinical practice.