An automatic interpretable deep learning pipeline for accurate Parkinson's disease diagnosis using quantitative susceptibility mapping and T1‐weighted images
Yida Wang,Naying He,Chunyan Zhang,Youmin Zhang,Chenglong Wang,Pei Huang,Zhijia Jin,Yan Li,Zenghui Cheng,Yu Liu,Xinhui Wang,Chen Chen,Jingliang Cheng,Fangtao Liu,Ewart Mark Haacke,Shengdi Chen,Guang Yang,Fuhua Yan
DOI: https://doi.org/10.1002/hbm.26399
IF: 4.8
2023-06-20
Human Brain Mapping
Abstract:In this study, we present a cascade deep learning pipeline for Parkinson's disease (PD) diagnosis with magnetic resonance images. Five brain nuclei regions were analyzed by our segmentation model before being used as a priori information by an anatomical attention‐guided convolutional neural network model for PD diagnosis. The proposed PD diagnosis model achieved area under the curves of 0.901 and 0.845 on independent internal and external testing cohorts, respectively. Parkinson's disease (PD) diagnosis based on magnetic resonance imaging (MRI) is still challenging clinically. Quantitative susceptibility maps (QSM) can potentially provide underlying pathophysiological information by detecting the iron distribution in deep gray matter (DGM) nuclei. We hypothesized that deep learning (DL) could be used to automatically segment all DGM nuclei and use relevant features for a better differentiation between PD and healthy controls (HC). In this study, we proposed a DL‐based pipeline for automatic PD diagnosis based on QSM and T1‐weighted (T1W) images. This consists of (1) a convolutional neural network model integrated with multiple attention mechanisms which simultaneously segments caudate nucleus, globus pallidus, putamen, red nucleus, and substantia nigra from QSM and T1W images, and (2) an SE‐ResNeXt50 model with an anatomical attention mechanism, which uses QSM data and the segmented nuclei to distinguish PD from HC. The mean dice values for segmentation of the five DGM nuclei are all >0.83 in the internal testing cohort, suggesting that the model could segment brain nuclei accurately. The proposed PD diagnosis model achieved area under the the receiver operating characteristic curve (AUCs) of 0.901 and 0.845 on independent internal and external testing cohorts, respectively. Gradient‐weighted class activation mapping (Grad‐CAM) heatmaps were used to identify contributing nuclei for PD diagnosis on patient level. In conclusion, the proposed approach can potentially be used as an automatic, explainable pipeline for PD diagnosis in a clinical setting.
radiology, nuclear medicine & medical imaging,neurosciences,neuroimaging