Assessment of the potential of convolutional neuronal networks in the differential diagnosis of Parkinson’s disease based on brain imaging [123I]FP-CIT SPECT

Aim/Introduction: To evaluate the potential of convolutional neural networks (CNN) in the differential diagnosis of Parkinson’s disease (PD) based on [123I]FP-CIT single-photon emission computed tomography (SPECT) images, compared to other machine learning-based classifiers. Materials and Methods: This work included 806 [123I]FP-CIT SPECT brain images (208 health controls and 598 with PD). Data were obtained from the Parkinson’s Progression Markers Initiative (PPMI) database (www.ppmi-info.org/data). For each subject, only the first scan was considered (baseline or screening). The protocol of image acquisition and pre-processing is available at http://www.ppmi-info.org/study-design/research-documentsand-sops/. CNN was compared against k-nearest neighbor (kNN), logistic regression (LG), decision trees (DT), support vector machines (SVM), and artificial neural networks (ANN) classifiers. The CNN classifier was trained with 2-dimensional image patches (dimensions: 88 mm x 82 mm) containing the striatal region, extracted from the head superior-inferior maximum intensity projection. The remaining classifiers were trained with five features extracted from the 3-dimensional striatal region: caudate binding potential, putamen binding potential, putamen to caudate ratio, the volume of the striatal region with “normal uptake”, and the major axis of that region. The minimum values extracted from each cerebral hemisphere were used. The split ratio of the dataset was 75:25 (75% for training and 25% for testing). Each of the five features was also considered individually to assess its potential for classification in terms of performance (accuracy, sensitivity, and specificity). Results: In the test dataset, the accuracy, sensitivity, and specificity of the CNN were 96%, 98%, and 91%, respectively. This finding was very similar to what we obtained with the other classifiers (kNN: 95%, 99%, 85%; LG: 94%, 97%, 86%, DT: 94%, 97%, 84%, SVM: 94%, 98%, 88%, and ANN: 94%, 97%, 86%). The accuracy differences are not statistically significant (Cochran Q test, p = 0.592). Individually, the feature that best differentiates PD from normal scans was the putamen binding potential with 93% accuracy, 93% sensitivity, and 94% specificity in the test dataset, based on the optimal cutoff (1.716) that maximizes Younden’s coefficient in the training dataset. Conclusion: CNN classifier proved to be as robust and accurate as the other classifiers frequently used in the type of problems as in this work, with the great advantage of using images as direct input. All machine learning-based classifiers tested are robust and very accurate in the classification of brain [123I]FP-CIT SPECT scans. Standard visual clinical evaluation should be complemented with quantification classification used also as a training tool.