Objective: The aim of this study is to implement and evaluate CNN-based DL approach for the     diagnosis of IPD. We trained CNNs on a large data set of high resolution nigrsome 1     MRI images and compared diagnostic performance with human specialist.

Background: Recently nigrosome 1 on susceptibility-weighted imaging (SWI) at 3T or 7T was     suggested as a new imaging biomarker of IPD. A recent meta-analysis showed that     overall sensitivity and specificity of nigrosome 1 MRI imaging for IPD versus normal     were 94.6% and 94.4%, respectively.

Method: We enrolled 570 subjects who visited our movement disorder clinic from November of 2014 to August of 2018. About 32,771 images of susceptibility map weighted imaging (SMWI), which was proposed to improve both contrast-to-noise ratios and signal-to-noise ratios for nigrosome 1 of SWI by using high spatial-resolution quantitative susceptibility mapping (QSM), were obtained from 570 subjects that were composed with 344 patients with Parkinson’s disease and 226 normal subjects at Gachon University Gil Medical Center and were used for learning, validation, and testing.

Results: The receiver operating characteristic (ROC) curves for 5 algorithms produced using 5-fold cross-validation, each of which was evaluated on the independent test data set (mean [SD] MR images of 176 subjects, 97.3 [0.52]). The mean (SD) values of the 5 areas under the ROC curve (AUC) were 0.9974 (0.0006). The diagnostic accuracy is 96.93±0.57 %. Sensitivity and specificity of the best performing developed algorithm for diagnosis of IPD were 97.33% and 100%, respectively.

Conclusion: In conclusion, the IPD diagnosis system using nigrosome 1 SMWI MRI images was found to have similar diagnostic performance with highly trained neuroradiologist and was verified in independent external dataset.

« Back to 2019 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/automated-diagnosis-of-idiopathic-parkinsons-disease-using-deep-convolutional-neural-networks/