PARNet: Deep neural network for the diagnosis of parkinson's disease

In this study, the successful network architecture we developed from scratch to diagnose COVID-19 has been retrained, using single photon emission computed tomography (SPECT) images to detect Parkinson's disease (PD). We aim to investigate whether a network trained on medical images can be adapted for the diagnosis of another disease successfully. This retrained neural network, PARNet, can detect PD patients. In this study, we use 1213 SPECT images as a dataset. The number of PD and healthy control (HC) group images is 1000 and 213, respectively. We divided the dataset into training (70%), validation (10%), and test (20%) sets. Our network shows outstanding performance with an accuracy of 95.43%, a sensitivity of 95.25%, a specificity of 95.70%, a precision of 97%, and an f1-score of 96%. Our method has the potential to improve the diagnosis and treatment of PD. PARNet, with high diagnosis performance, can contribute to assisting clinicians in diagnosing PD at an earlier. PARNet network based on COV19-ResNet architecture showed performance similar to even high to that of the larger pre-trained models of ImageNet in diagnosing PD. This network can be easily retrained with images from different medical domains to detect various diseases.

Automated summary

This study demonstrates the successful retraining of a network architecture, originally developed for COVID-19 diagnosis, for the detection of Parkinson's disease using SPECT images. The retrained neural network, PARNet, shows outstanding performance in diagnosing PD and has the potential to improve the diagnosis and treatment of the disease.