Research on Diffusion Kurtosis Imaging of the Brain Based on Deep Learning

Precise evaluation of the deviation of water molecule diffusion in the brain tissue from the Gaussian diffusion model is crucial for diagnosing brain disorders. Diffusion kurtosis imaging (DKI) is a technique that may accurately describe the non-Gaussian diffusion characteristics of water molecules in brain tissue. The traditional method for achieving diffusion kurtosis imaging is constrained weighted linear least squares estimation. However, the traditional constrained weighted linear least squares estimation method has greatly reduced its clinical practicability due to its excessively long data acquisition time. Consequently, we propose using a deep learning-based approach instead of the traditional constrained weighted linear least-squares estimation method. We suggest utilizing the DKI-Net model, which relies solely on 2D convolution, and the DKI-Transformer model, which has an improved self-attention mechanism module, to estimate DKI-derived metrics. DKI-Net with residual properties extracts accurate and rich blocks of feature information and highly estimates voxels. The DKI-Transformer model can extract global voxel correlation characteristics, the estimation results have a high peak signal-to-noise ratio and a high structural similarity index compared to the reference labeling and exhibit distinct boundaries of microscopic features. Compared with traditional methods, these methods utilize the strong fitting capabilities of deep learning to estimate high-quality rotationally invariant scalar measures using a small number of DWI data points, significantly decreasing the number of scans and the duration of scanning exponentially. This approach provides a perspective that promises to address the problem of long imaging times in diffusion kurtosis imaging (DKI), facilitating the resolution of the clinical limitations of DKI. Consequently, it is anticipated that the range of applications of DKI will broaden.