Multivariable model for diagnosis of neurodegenerative diseases using I-123 Ioflupane incorporating machine learning

Introduction: Dopamine transporter imaging with ¹²³I-ioflupane (DaTSCAN) has been used in clinical practice to differentiate Parkinson syndrome/diseases and Lewy body diseases among various movement disorders and dementia. The abnormality of single-photon emission computed tomography (SPECT) images has been interpreted visually combined with region-of-interest (ROI)-based semiquantitative uptake indices such as specific binding ratio (SBR). However, the accurate diagnosis based on such ROI-based indices is still challenging due to indeterminate borderline results. To enhance the diagnostic ability without setting ROI, we aimed to create a model using machine-learning features and multivariable models.

Methods: A total of 239 patients were consecutively selected in two hospitals, and two groups of patient databases, i. e., training (n=137) and validation (n=102) datasets, were created. First, the training database consisted of 137 consecutive patients suspected of having Parkinson syndrome/diseases and Lewy body diseases (PSD/Lewy) in one hospital, and average age was 75 ± 8 years (51 - 92 years, male 64%) including PSD/Lewy (77%) and non-PSD/Lewy (23%) groups. The transaxial DICOM images were preprocessed to select slices of the highest count on the striatum, and the contour of the head was trimmed automatically for subsequent data processing. For feature recognition, well-trained nuclear medicine specialists classified images into four features; namely, (F1) high or low uptake, (F2) symmetric or asymmetric striatal uptake, (F3) uptake in caudate and putamen of comma-like (normal caudate and putaminal uptake) or dot-like (relatively low putaminal uptake) appearance, and (F4) normal or abnormal uptake overall. Machine learning algorithms tested were logistic regression (LR), k-nearest neighbors (kNN), and gradient boosted trees (GBT) methods. The datasets were trained and validated using four-fold cross validation (75% and 25% for training and test, respectively). Appropriate machine-learning classifiers were determined with receiver-operating characteristics (ROC) analyses. Second, the validation dataset from the other hospital was used, and the neurological diagnosis was made independently by neurology experts, and classified into PSD/Lewy and non-PSD/Lewy groups. The three diagnostic models were compared: Model 1, ROI-based calculation of SBR and asymmetry index (average count ratio of laterality); Model 2, machine-learning based judgement of abnormality (F4); Model 3, multivariable model using age and probabilities of three features (F1, F2, and F3). The diagnostic accuracy was compared among three models using ROC area under the curve analysis (AUC).

Results: Based on the training database with four-fold cross validation, ROC AUC was high with the high/low feature by all methods with LR, kNN, GBT, and ROI-based SBR (AUC 0.92-0.96). With respect to the symmetric/asymmetric feature, kNN method showed the highest AUC of 0.68, followed by GBT (0.67), ROI-based asymmetry index (0.64), and LR (0.58) methods. The comma/dot-like feature showed that AUC of 0.92 for LR, 0.88 for GBT, 0.79 for kNN, and 0.81 by ROI-based putamen/caudate average count ratio. Based on the validation dataset, the diagnostic metrics of PSD/Lewy profile measured by AUC (± standard error) were 0.86 ± 0.04, 0.88 ± 0.04, and 0.93 ± 0.02 for Model 1 (ROI-based), Model 2 (machine learning-based), and Model 3 (machine-learning based three features + age), respectively. The difference in AUC was significant between Models 3 versus1 (p = 0.027) and Models 3 versus 2 (p = 0.029), and AUC of Model 3 was best.

Conclusions: For the image interpretation of ¹²³I-ioflupane, machine learning can be used to determine image features such as high/low, symmetric/asymmetric, and comma/dot-like profiles. The combined multivariate model with three features plus age showed the highest diagnostic accuracy for differentiating PSD/Lewy group compared with the conventional ROI-based method.

• Download figure
• Open in new tab
• Download powerpoint