Artificial intelligence in the diagnosis of Parkinson's disease from ioflupane-123 single-photon emission computed tomography dopamine transporter scans using transfer learning.

Abstract

The objective of this study was to identify the extent to which artificial intelligence could be used in the diagnosis of Parkinson's disease from ioflupane-123 (¹²³I) single-photon emission computed tomography (SPECT) dopamine transporter scans using transfer learning. A data set of 54 normal and 54 abnormal ¹²³I SPECT scans was amplified 44-fold using a process of image augmentation. This resulted in a training set of 2376 normal and 2376 abnormal images. This was used to retrain the top layer of the Inception v3 network. The resulting neural network functioned as a classifier for new ¹²³I SPECT scans as either normal or abnormal. A completely separate set of 45 ¹²³I SPECT scans were used for final testing of the network. The area under the receiver-operator curve in final testing was 0.87. This corresponded to a test sensitivity of 96.3%, a specificity of 66.7%, a positive predictive value of 81.3% and a negative predictive value of 92.3%, using an optimum diagnostic threshold. This study has provided proof of concept for the use of transfer learning, from convolutional neural networks pretrained on nonmedical images, for the interpretation of ¹²³I SPECT scans. This has been shown to be possible in this study even with a very small sample size. This technique is likely to be applicable to many areas of diagnostic imaging.