DeepDTI: High-Fidelity Six-Direction Diffusion Tensor Imaging Using Deep Learning

Computer Science Information Science Artificial Intelligence Neurology Biophysics Medical Imaging Life Sciences Medicine
diffusion tensor imaging diffusion tractography tract-specific analysis deep learning residual learning convolutional neural network data redundancy

DeepDTI: High-Fidelity Six-Direction Diffusion Tensor Imaging Using Deep Learning

Research Background and Motivation

Detail Restoration of Anatomical Information by DeepDTI Diffusion Tensor Imaging (DTI) boasts unparalleled advantages in mapping the microstructure and structural connectivity of live human brain tissue. However, traditional DTI techniques require extensive angular sampling, leading to prolonged scanning times, which limits their application in routine clinical practice and large-scale studies. To overcome this bottleneck, researchers have developed a novel DTI processing framework called DeepDTI, which minimizes DTI data requirements through data-driven supervised deep learning. This paper aims to demonstrate how DeepDTI significantly reduces the amount of DTI sampling data, thereby achieving faster scan speeds while maintaining high-quality imaging results.

Source of the Paper

The primary authors of this paper include Qiyuan Tian, Berkin Bilgic, Qiuyun Fan, Congyu Liao, and others. They are affiliated with the Athinoula A. Martinos Center for Biomedical Imaging at Massachusetts General Hospital, Harvard Medical School, the Department of Health Sciences and Technology at MIT, and the Department of Electrical Engineering at Stanford University. The paper was published in the journal Neuroimage on October 1, 2020.

Research Process

Research Methods

The design of the DeepDTI framework is based on a deep understanding of the physics of diffusion MRI. Its innovation lies in utilizing a 10-layer three-dimensional Convolutional Neural Network (CNN) to map the input non-diffusion-weighted (b=0) image and six optimized diffusion-weighted images (DWI), along with T1-weighted and T2-weighted images, to the residual between the input images and the output high-quality images. The input and output of DeepDTI are specially designed to improve CNN performance through residual learning and achieve high-quality DWI tensor fitting to generate directional DTI metrics for tractography.

Data Processing:

  1. Data Collection:

    • This study used preprocessed diffusion, T1-weighted, and T2-weighted MRI data of 70 unrelated healthy subjects from the Human Connectome Project (HCP) public database.
    • The diffusion MRI data were acquired at a 1.25 mm isotropic resolution, including four b-values (0, 1, 2, 3 ms/μm²) and two phase encoding directions (left-right and right-left).

  2. Image Preprocessing:

    • The images underwent susceptibility-induced distortion and eddy current distortion correction using FSL software. Brain tissue segmentation was performed on the acquired image data, excluding cerebrospinal fluid (CSF).
    • The raw data were fitted with diffusion tensors using the least squares method to obtain a series of DTI metrics.

  3. Model Training:

    • The input images for the CNN were standardized, and the CNN parameters were optimized using the Adam optimizer. Data sets for training and validation were generated by randomly selecting raw data from different angular samplings.
    • Model performance was evaluated using metrics such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM).

Experimental Results:

The output images of DeepDTI showed significantly superior noise suppression and detail fidelity compared to the raw images and the latest BM4D denoising algorithm. Quantitative evaluation results indicated significant improvements in PSNR and SSIM for DeepDTI output images. Key results include: - The PSNRs for input and output b=0 images and DWI images were 34.6 dB and 31.9 dB, respectively. - For 20 evaluation subjects, the mean squared error of DTI metrics generated by DeepDTI was significantly lower than that of raw data and BM4D denoised data. - The core distance of the primary white matter fiber bundles in the tractography results generated by DeepDTI was 1-1.5 mm compared to the ground truth results generated using all data (18 b=0 images and 90 DWIs).

Research Conclusions

DeepDTI significantly reduces the DTI sampling requirements while maintaining or even surpassing the imaging quality and DTI analysis performance of existing methods. By reducing the acquisition requirements to one b=0 image and six DWI images, DeepDTI shortens DTI scan times to 30-60 seconds, making it more practical and usable in clinical and research settings. With DeepDTI, high-fidelity DTI can become a routine imaging modality, greatly improving scanning efficiency, especially suitable for patients who have difficulty remaining still or for young children.

Research Highlights

  • The innovative DeepDTI framework design improves CNN performance through residual learning, achieving high-quality diffusion tensor imaging.
  • With the minimum sampling data (one b=0 image and six DWI), DeepDTI significantly outperforms existing methods in terms of fidelity and accuracy.
  • The application of deep learning makes complex microstructural imaging analysis possible, enhancing the utilization and accuracy of diffusion MR data.

Significance and Application Value

DeepDTI’s success showcases the immense potential of deep learning in uncovering hidden information within imaging data. This method not only improves the efficiency and quality of DTI imaging but also offers broad prospects for further research and application, especially in clinical and neuroscience research where fast and high-quality DTI is required.

Through this study, the DTI imaging time is greatly shortened, making it a part of routine imaging examinations, enhancing clinical diagnosis and research efficiency, and providing significant practical value, especially when dealing with unsteady patients and young children.