Examination of simple artificial intelligence-based analysis of dopamine transporter scintigraphy for supporting a diagnosis of Parkinson's disease.
Authors
Affiliations (9)
Affiliations (9)
- Department of Neurology, Nagoya University Graduate School of Medicine, 65 Tsurumai-Cho Showa-Ku, Nagoya, Aichi, 466-8550, Japan.
- Department of Neurology, NHO Higashinagoya National Hospital, 5-101 Umemorizaka Meito-Ku, Nagoya, Aichi, 465-8620, Japan.
- Department of Neurology, Japanese Red Cross Aichi Medical Center Nagoya Daini Hospital, 2-9 Myoken-cho Showa-ku, Nagoya, Aichi, 466-8650, Japan.
- Innovative Research Center for Preventive Medical Engineering, Institute of Innovation for Future Society, Nagoya University, Furo-cho Chikusa-Ku, Nagoya, Aichi, 464-8601, Japan.
- Brain and Mind Research Center, Nagoya University, 65 Tsurumai-Cho Showa-Ku, Nagoya, Aichi, 466-8550, Japan.
- Department of Neurology, TOYOTA Memorial Hospital, 1-1 Heiwa-cho, Toyota, Aichi, 471-8513, Japan.
- Department of Radiological Technology, Nagoya University Hospital, 65 Tsurumai-Cho Showa-Ku, Nagoya, Aichi, 466-8550, Japan.
- Functional Medical Imaging, Biomedical Imaging Sciences, Division of Advanced Information Health Sciences, Department of Integrated Health Sciences, Nagoya University Graduate School of Medicine, 1-1-20 Daikominami Higashi-Ku, Aichi, 461-8673, Nagoya, Japan.
- Department of Neurology, Nagoya University Graduate School of Medicine, 65 Tsurumai-Cho Showa-Ku, Nagoya, Aichi, 466-8550, Japan. [email protected].
Abstract
In the Movement Disorder Society criteria for the diagnosis of Parkinson's disease (PD), evaluation of the presynaptic dopamine system should be performed using dopamine transporter single-photon emission computed tomography (DAT SPECT). However, it is difficult for unexperienced physicians to detect a mild defect. Here, we attempted to develop a simple deep learning-based image analysis method to evaluate DAT SPECT images. We used data from 300 patients who were diagnosed with PD and 102 control patients with non-neurodegenerative diseases as the artificial intelligence (AI) development cohort. For validation, we analyzed the data of 96 patients with PD from an independent cohort. We divided the development cohort into the training and test sets. Using the training set, we performed transfer learning using six pre-trained convolutional neural network architectures, and created AI models. We evaluated their accuracy, sensitivity, and area under the receiver operating characteristic curve, and further confirmed their validity by using the validation cohort. In addition, we compared the accuracy of the best AI model with that of two experienced neurologists and a resident. The selected AI model could interpret DAT SPECT images with an accuracy of 0.959; accuracy in the validation cohort was 0.8959-1. There was no significant difference between the accuracy of the AI model and physicians. Our simple AI model for the interpretation of DAT SPECT images was accurate and robust. Its accuracy was equivalent to that of physicians.