Naying He1, Yu Liu1, Bin Xiao2, Junchen Li3, Chencheng Zhang4, Yijie Lai4, Feng Shi5, Dinggang Shen5, Yan Li1, Hongjiang Wei6, Ewart Mark Haacke1,7, Weibo Chen8, Qian Wang2, Dianyou Li4, and Fuhua Yan1
1Radiology, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China, 2Institute for Medical Imaging Technology, School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China, Shanghai, China, 3Changshu Hospital Affiliated to Nanjing University of Chinese Medicine, No. 6 Huanghe Road, Changshu, China, Changshu, China, 4Department of Neurosurgery, Center for Functional Neurosurgery, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China, Shanghai, China, 5Shanghai United Imaging Intelligence Co., Ltd., Shanghai, China, Shanghai, China, 6Institute for Medical Imaging Technology, Department of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China, Shanghai, China, 7Department of Radiology, Wayne State University, Detroit, Michigan, USA, Detroit, MI, United States, 8Philips Healthcare，Shanghai，China, Shanghai, China
1Radiology, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China, 2Institute for Medical Imaging Technology, School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China, Shanghai, China, 3Changshu Hospital Affiliated to Nanjing University of Chinese Medicine, No. 6 Huanghe Road, Changshu, China, Changshu, China, 4Department of Neurosurgery, Center for Functional Neurosurgery, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China, Shanghai, China, 5Shanghai United Imaging Intelligence Co., Ltd., Shanghai, China, Shanghai, China, 6Institute for Medical Imaging Technology, Department of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China, Shanghai, China, 7Department of Radiology, Wayne State University, Detroit, Michigan, USA, Detroit, MI, United States, 8Philips Healthcare，Shanghai，China, Shanghai, China
This QSM based radiomics model performed
best with an AUC of 0.897 to predict the STN-DBS motor outcome in PD. In addition, the
threshold probability of the RA-ML model can differentiate surgical responders
and non-responders.
Fig. 2. Graph shows receiver
operating characteristic curve to assess the utility of two different models
with the clinical variables included or not (with and w/o clinical information)
for predicting global motor outcome.
Fig. 1. Illustration of the processing
pipeline of the radiomics model with
machine learning (RA-ML). RF=radiomics feature; RFE= recursive feature
elimination