Application of Pseudo-Three-Dimensional Residual Network to Classify the Stages of Moyamoya Disease

Abstract

It is essential to assess the condition of moyamoya disease (MMD) patients accurately and promptly to prevent MMD from endangering their lives. A Pseudo-Three-Dimensional Residual Network (P3D ResNet) was proposed to process spatial and temporal information, which was implemented in the identification of MMD stages. Digital Subtraction Angiography (DSA) sequences were split into mild, moderate and severe stages in accordance with the progression of MMD, and divided into a training set, a verification set, and a test set with a ratio of 6:2:2 after data enhancement. The features of the DSA images were processed using decoupled three-dimensional (3D) convolution. To increase the receptive field and preserve the features of the vessels, decoupled 3D dilated convolutions that are equivalent to two-dimensional dilated convolutions, plus one-dimensional dilated convolution, were utilized in the spatial and temporal domains, respectively. Then, they were coupled in serial, parallel, and serial–parallel modes to form P3D modules based on the structure of the residual unit. The three kinds of module were placed in a proper sequence to create the complete P3D ResNet. The experimental results demonstrate that the accuracy of P3D ResNet can reach 95.78% with appropriate parameter quantities, making it easy to implement in a clinical setting.