RISNet: A variable multi-modal image feature fusion adversarial neural network for generating specific dMRI images

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Medical Image Analysis

Background:

• Diffusion magnetic resonance imaging (dMRI) b-values influence image contrast and calculation accuracy.
• Imbalance in lower and higher b-value macaque dMRI data hinders computational neuroscience.
• Existing generative adversarial networks struggle with multi-center, small-sample macaque brain datasets.

Purpose of the Study:

• To address the scarcity of lower b-value dMRI data in macaques.
• To improve the accuracy of computational neuroscience analyses using macaque brain imaging.
• To develop a robust medical image conversion method for macaque dMRI.

Main Methods:

• Proposed RISNet, a variable multi-modal image feature fusion adversarial neural network.
• Introduced Rapid Insertion Structural (RIS) to enhance model generalization by fusing multi-modal features.
• Utilized T1 and higher b-value brain images as inputs to generate lower b-value images.

Main Results:

• Achieved an average improvement of 1.8211 in PSNR and 0.0111 in SSIM compared to existing methods.
• Demonstrated sound visual effects in qualitative observations.
• Showcased strong generalization ability in Diffusion Tensor Imaging (DTI) estimation.

Conclusions:

• RISNet effectively solves the dMRI brain image conversion problem in macaques.
• The method enhances the quality and utility of macaque dMRI data.
• Provides strong support for future neuroscience research utilizing macaque models.