%0 Journal Article %A Matthew Adams %A Jing Tang %T Convolutional neural network based motor score prediction using DAT SPECT imaging of Parkinson’s disease %D 2019 %J Journal of Nuclear Medicine %P 405-405 %V 60 %N supplement 1 %X 405Objectives: Dopamine transporter (DAT) SPECT imaging has seen wide use in the diagnosis of Parkinson’s disease (PD). Our previous efforts found correlations between clinical assessments and DAT SPECT radiomic features. The goal of this study is to predict clinical motor function evaluation scores using the longitudinal DAT SPECT images and non-imaging features. This is approached by developing a convolutional neural network (CNN) based prediction technique. Methods: Data of 252 subjects from the Parkinson’s Progression Markers Initiative (PPMI) database were included in this study. Outcome was set as the motor part (III) score of the unified Parkinson’s disease rating scale (UPDRS) at year 4. The prediction input data consisted of the DAT SPECT images and UPDRS_III scores acquired at year 0 and year 1. Before the prediction process, the required data was gathered and pre-processed. The UPDRS_III scores were taken into consideration when a patient was not on medication during evaluation. The score variability was mitigated by averaging each score with any available scores 6 months before or after. All DAT SPECT voxel intensities were normalized to range from 0 to 1, corresponding to the minimum and maximum voxel values for a given image respectively. The data was augmented by including a left/right-flipped copy of each image. A CNN was developed in Tensorflow using the specified inputs and outputs to train and test the network. Each image connects to its own hidden layers that use rectified linear units as activation functions. The hidden layers include three 3-D convolutional layers of sizes 7×7×7, 5×5×5, and 3×3×3 with stride lengths of three, one, and one respectively. Each convolutional layer is followed by a max-pooling layer of size 3×3×3 with a stride length of two, which reduces the data size and the required computations. Three fully connected layers with 1024, 256, and 64 neurons subsequently connect the convolutional layers to the output layer. During training, the Adam optimization algorithm was used to update the network parameters, with the mean square error as the cost function. The learning rate was initialized to 0.001 and the network was trained for a total of 25 iterations. The fully connected layer includes a 0.5 dropout, which was only used for training. The trained network was tested using 10-fold cross-validation and the performance was evaluated based on the absolute difference between a predicted score and its actual score correspondence. The evaluation was performed for prediction with and without DAT images, the latter implemented by setting unused inputs to zero. A two-sample t-test was used to determine whether there was statistically significant difference in prediction of the two cases. Results: Using the longitudinal DAT SPECT images and UPDRS_III scores for prediction resulted in an average difference of 6.51±5.17 between the predicted and actual scores. In contrast, voiding all image inputs results in an average difference of 7.93±6.02. The two-sample t-test to compare the two cases results in a two-tailed p-value of 0.005, demonstrating the added value of the image data. Conclusions: With the non-imaging feature UPDRS_III and the DAT SPECT images from year 0 and year 1, we predicted the year 4 UPDRS_III motor function score using the developed CNN scheme. It was demonstrated that addition of DAT SPECT images to non-imaging features significantly improved the result for this challenging task, without requiring segmentation and feature extraction. We expect that continuing efforts will further improve diagnostic accuracy and outcome prediction in PD. %U