Artificial neural network based outcome prediction in DAT SPECT imaging of Parkinson's Disease

Objectives: Dopamine transporter (DAT) SPECT imaging has been widely used in the diagnosis of Parkinson’s disease (PD). Compared to visual interpretation of DAT SPECT images, quantitative analysis has the significant potential to enable early stage detection of disease and improve progression tracking. Our past radiomic efforts focused on seeking the correlations between DAT SPECT radiomic measures and clinical assessments. By contrast, the goal of this study is to predict clinical outcome using DAT SPECT radiomic features and non-imaging features at baseline (year 0). We accomplish this goal through developing an artificial neural network (ANN) based prediction technique.

Methods: Sixty-nine subjects in the Parkinson’s Progression Markers Initiative database were included in our study. The subjects were imaged by SPECT 4+-0.5 h following injection of 111-185 MBq of 123I-Ioflupane, and also underwent high-resolution 3T MRI scans. The MR images were segmented to obtain the structures of caudate, putamen, and occipital cortex which were then overlaid on the co-registered SPECT images. Outcome was set as motor part (III) of the unified Parkinson’s disease rating scale (UPDRS) at year 4. From the more affected caudate in the SPECT image at baseline, we extracted a total of 92 image features, including 14 first-order intensity features, 21 shape features, and 57 second- and higher-order textural features. Six non-imaging features were also included in the study, i.e. gender, age, disease duration (DD)_sympt, DD_diag, UPDRSIII, and Montreal cognitive assessment (MoCA). We performed a t-test for each of the 98 features to classify the patients to group 1 (UPDRSIII of year 4 < 30) and group 2 (UPDRSIII of year 4 > =30). Through sorting the p-values from the t-tests in ascending order, we selected the top 10 features. The mean standardized uptake value (SUV mean), as a conventional feature, was also used to predict outcome at year 4 for a comparison purpose. Using the 10 selected features and the SUV mean at baseline as the inputs, we applied ANNs with one hidden layer to predict which group each patient belonged to. The number of neurons in the hidden layer was set to 1, 3, and 5, respectively. Having limited number of patients, we assessed the prediction accuracy of different individual features and different ANN architectures via 6900 rounds of leave-one-out cross-validation. In each round, one subject was left for testing and the remaining subjects for training. The means and standard deviations for prediction accuracy, specificity, and sensitivity were calculated from the rounds. For every feature used, the final prediction result was optimized from the three ANNs of different neuron numbers.

Results: After sorting the t-test p-values, 1 non-imaging feature (UPDRSIII at baseline), 4 shape features, and 5 textual features were selected as the top 10 features. Among the selected individual features and the SUV mean, the highest prediction accuracy of 70.9% was achieved by the textural feature HGZE_GLSZM, and the second highest accuracy of 70.7% by the UPDRSIII at baseline. By contrast, the conventional feature SUV mean gave the prediction accuracy of 61.8% and the sensitivity of 3.7%, meaning that it barely recognized subjects that belonged to group 2. Among all the image features, the HGZE_GLSZM showed performance superior to all the intensity and shape features, especially the conventional feature, in predicting outcome.

Conclusion: Using the non-imaging feature UPDRSIII and the imaging features extracted from the DAT SPECT images at baseline, we predicted year 4 outcome using the developed ANN scheme. In doing so, we established an accuracy of 70% for this challenging task and demonstrated the feasibility of applying ANNs on textural and non-imaging features for clinical prediction. We expect that continuing efforts will augment diagnostic accuracy and improve outcome prediction in PD. Research Support: NSF ECCS-1454552 and M. J. Fox Foundation ID 9036.01

• Download figure
• Open in new tab
• Download powerpoint