RT Journal Article
SR Electronic
T1 <strong>A deep learning-based approach for disease detection in the projection space of DAT-SPECT</strong><strong> images of patients with Parkinson’s disease</strong>
JF Journal of Nuclear Medicine
JO J Nucl Med
FD Society of Nuclear Medicine
SP 509
OP 509
VO 61
IS supplement 1
A1 Kevin Leung
A1 Wenyi Shao
A1 Lilja Solnes
A1 Steven Rowe
A1 Martin Pomper
A1 Yong Du
YR 2020
UL http://jnm.snmjournals.org/content/61/supplement_1/509.abstract
AB 509Objectives: Dopamine transporter (DAT) single photon emission computed tomography (SPECT) imaging is an established tool that is routinely used for diagnostic purposes in Parkinson’s disease (PD) (1). While visual analysis of DAT-SPECT images with striatal DAT quantification is standard practice in clinical settings, visual analysis is also suboptimal due to interobserver variability (2). DAT-SPECT images reconstructed by iterative methods, such as ordered subsets expectation maximization (OSEM), are strongly affected by reconstruction parameters (3). Additionally, heavy post reconstruction smoothing is often applied in the clinic to reduce noise (3). Due to all those processes, some information present in the original projection data could be lost in reconstructed images. We hypothesize that there may be more information present in the projection data that is important for the disease detection task. Recently, deep learning-based methods have shown promise in a wide range of medical image analysis tasks including disease detection and classification tasks (4,5). In this work, we aimed to develop a deep learning-based method for detecting patients with Parkinson’s disease from normal subjects using the projection data from DAT-SPECT. Methods: A total of 659 3D DAT-SPECT images of healthy controls (HCs) and patients with PD were used to train the proposed method. This data was extracted from the open-source Parkinson’s Progression Markers Initiative (PPMI) database (6). Projection data were generated from the 659 DAT-SPECT image volumes using an analytical projector that incorporated attenuation and spatially varying resolution. The 659 subjects were randomly partitioned into training, validation and test datasets containing 527 (168 HCs and 359 PD), 65 (21 HCs and 44 PD) and 67 (21 HCs and 46 PD) subjects, respectively, using an 80%/10%/10% split. A deep 3D convolutional neural network (3D-CNN) was developed to classify PD patients from HCs (Fig. 1a). The proposed network architecture consisted of a series of 3D strided convolutional layers followed by a global average pooling layer and a fully connected layer. The 3D projection data were used as input to the 3D-CNN. The output of the network was the predicted likelihood that the subject was a healthy control or PD patient. The network was trained and optimized with a class-weighted cross-entropy loss function and a stochastic gradient-based optimization algorithm, Adam (7). The network was trained with early stopping based on monitoring the error on the validation set to regularize the network and prevent overfitting (8). The proposed method was then evaluated on the test set. The proposed method was evaluated by assessing standard evaluation metrics, including the overall classification accuracy, true positive fraction (TPF), true negative fraction (TNF), area under receiver operating characteristic curve (AUROC) and confusion matrix. Overall accuracy was defined as the number of correctly classified observations divided by the total number of observations. Receiver operating characteristic (ROC) analysis was performed on the output of the proposed method on the test set. Results: The proposed method yielded an overall accuracy of 97.0% (95% confidence interval (CI): 92.8%, 100.0%)), a TPF of 0.98, a TNF of 0.95 and an AUROC of 0.96 on the test set. The ROC curve and confusion matrix are shown in Fig. 1b and c, respectively, for the proposed method on the test set. Conclusions: A deep learning-based approach for detection of subjects with PD in DAT-SPECT 3D projection data was developed and showed significant promise towards classification of subjects with PD. For future work, we will compare performance of this approach with networks that classify the reconstructed DAT-SPECT images.