Deep Learning with Anatomical Attention Mechanism for Distinguishing Parkinson’s Disease from Normal Controls in MR imaging

We proposed an automatic cascaded framework based on deep learning to segment deep brain nuclei and distinguish Parkinson’s disease from normal controls using quantitative susceptibility mapping (QSM) images. A 3D CA-Net model integrating channel attention, spatial attention and scale attention module was utilized to segment 5 brain nuclei from QSM and T1W data. Then, the QSM images and the segmented brain nuclei ROIs were fed into the SE-ResNeXt50 with anatomical attention mechanism to get the predicted PD probability. The proposed method provided good interpretability and achieved AUC values of 0.97 and 0.90 on training and testing cohort, respectively.