| Summary: | Diffusion kurtosis imaging (DKI) is an advanced diffusion imaging method that captures complex brain microstructural properties; however, it often has a lengthy acquisition time compared to conventional diffusion tensor imaging (DTI). Recently, a deep learning-based method has shown the potential for reducing the number of diffusion-weighted images (DWIs) required to compute the rotationally invariant scalar measures to twelve. In this study, we propose a three-dimensional (3D) convolutional neural network (CNN) to estimate the scalar measures. This network further improves the performance of the deep learning-based method with a largely reduced number of required DWIs. In our approach, all the DTI and DKI measures were estimated using a single network, and a hierarchical structure was introduced to customize the outputs based on their computational complexities and to learn the commonalities of the measures. Moreover, 3 × 3 × 3 convolution kernels were introduced to extract features from the 3D input patches and utilize the spatial context from adjacent neighborhoods, which also strengthened the network's robustness against noise. The proposed method was evaluated with two datasets. The results showed that, compared with the previous method that used an artificial neural network, our proposed hierarchical CNN provided enhanced efficiency for estimating all eight diffusion measures. It also improved the robustness against noise and retained the fine structures with only a few DWIs (as few as eight). This result suggests that it is possible to achieve kurtosis mapping in most clinical scanners within one minute, which could significantly extend the clinical utility of the DKI.
|
|---|
