Relevance of 18F-DOPA visual and semi-quantitative PET metrics for the diagnostic of Parkinson disease in clinical practice: a machine learning-based inference study

Abstract

Purpose: To decipher the relevance of visual and semi-quantitative 6-fluoro-(18F)-L-DOPA (¹⁸F-DOPA) interpretation methods for the diagnostic of idiopathic Parkinson disease (IPD) in hybrid positron emission tomography (PET) and magnetic resonance imaging.

Material and methods: A total of 110 consecutive patients (48 IPD and 62 controls) with 11 months of median clinical follow-up (reference standard) were included. A composite visual assessment from five independent nuclear imaging readers, together with striatal standard uptake value (SUV) to occipital SUV ratio, striatal gradients and putamen asymmetry-based semi-quantitative PET metrics automatically extracted used to train machine learning models to classify IPD versus controls. Using a ratio of 70/30 for training and testing sets, respectively, five classification models-k-NN, LogRegression, support vector machine, random forest and gradient boosting-were trained by using 100 times repeated nested cross-validation procedures. From the best model on average, the contribution of PET parameters was deciphered using the Shapley additive explanations method (SHAP). Cross-validated receiver operating characteristic curves (cv-ROC) of the most contributive PET parameters were finally estimated and compared.

Results: The best machine learning model (k-NN) provided final cv-ROC of 0.81. According to SHAP analyses, visual PET metric was the most important contributor to the model overall performance, followed by the minimum between left and right striatal to occipital SUV ratio. The 10-time cv-ROC curves of visual, min SUVr or both showed quite similar performance (mean area under the ROC of 0.81, 0.81 and 0.79, respectively, for visual, min SUVr or both).

Conclusion: Visual expert analysis remains the most relevant parameter to predict IPD diagnosis at 11 months of median clinical follow-up in ¹⁸F-FDOPA. The min SUV ratio appears interesting in the perspective of simple semi-automated diagnostic workflows.

Keywords: Fluorodopa F 18; Machine learning; Parkinson's disease; Positron-emission tomography.

Fig. 1

Flowchart of the patient selection process. From 367 patients, 110 were finally retained, including 48 IPD and 62 controls

Fig. 2

Image processing and data extraction. The whole dataset was assessed both visually and automatically. For visual analysis (A), five independent readers blindly re assessed all the ¹⁸F-DOPA PET data by using the international recommendation procedure guideline. An overall composite “visual” interpretation based on the five readers results was generated for each ¹⁸F-DOPA PET. For automated analysis (B), the ¹⁸F-DOPA PET/MRI data were processed in a dedicated neuroimaging pipeline (Free surfer) to be standardized and labeled. For this purpose, each T1w MRI was normalized into the MNI space, and the transformation was applied to the corresponding ¹⁸F-DOPA PET. Numerous metrics could be extracted automatically from the standardized PET data. Finally, a subset of targeted PET metrics was retained (C) based on their conceptual relevance and explored to identify potential high collinearity (D)

Fig. 3

Statistical analyses. After identifying potential collinearity between the PET metrics, the whole dataset (visual binary interpretation and semi-quantitative PET metrics from 110 ¹⁸F-DOPA PET/MRI) was split into training (70%) and test (30%) sets. Five machine learning classifiers (KNN, Log Regression-Log Reg, Support Vector Machine-SVM, Random Forest-RF and tree gradient boosting) were trained to predict the final diagnosis at 11 months of median follow-up (IPD or control) on the training set by using a nested k-fold cross-validation procedure. (Each model parameters are fine-tuned and cross-validated while optimizing the bias of over fitting.) The overall nested cross-validation procedure was repeated 100 times. The best model on average was applied on the test set to provide general unbiased accuracy. Finally, the contribution of each ¹⁸F-DOPA PET parameter (the visual and four semi-quantitative metrics) in the model predictions was deciphered

Fig. 4

Visual analyses of ¹⁸F-DOPA PET/MRI according to the international guidelines [11]. A This IPD patient, visually considered normal. The clinical follow-up and had symptoms relief after L-DOPA therapeutic test classified him as FN. Only min SUVr was below the 95% confidence interval estimated from controls. B This IPD presented unilateral right posterior putamen (white arrow) pre-synaptic dopamine denervation and global striatal ¹⁸F-DOPA decrease with faintly increased background unspecific activity. PET right putamen dopamine denervation was confirmed with symptoms relief after L-DOPA therapeutic test

Fig. 5

Subgroups distributions of ¹⁸F-DOPA PET semi-quantitative metrics. A fused ¹⁸F-DOPA PET/T1weighted MRI normalized in the MNI space of a control subject. B MNI-normalized T1-weighted MRI, overlaid with corresponding Freesurfer-based automated segmentation labels (C). Subgroups distribution of the four ¹⁸F-DOPA PET extracted metrics, for which a statistical between group difference was verified, values and p values are detailed in Table 3

Fig. 6

Heatmap of the between PETmetrics correlations (Spearman rank test)

Fig. 7

Shap plot of features’ contribution on the best classifier model output. A Stacked bars of absolute value of the SHAP values for each feature sorted by their importance across all patients for classification. “Class 0” corresponds to the controls and “Class 1” to the IPD group. For classification of IPD versus control, PET visual score is the most contributive parameter, then min SUV ratio is the second most contributive parameter. The three others do not appear relevant for the classification of IPD versus controls here. B SHAP distribution on x-axis of the five features for every patient and control. The impact is correlate to the absolute value of the SHAP value. The higher (redder) feature value, the more it contributes to IPD classification. On the opposite the lower (bluer) feature value, the more it contributes to control classification. Min SUV ratio and intra-striatal gradient are inversely correlated to IPD classification as represent dopamine denervation at different levels

Fig. 8

ROC curves comparison. Left panel: best model AUC (tenfold cross-validated) with visual PET metric alone; middle panel: best model AUC (tenfold cross-validated) with min SUVr PET metric alone; and right panel: best model AUC (tenfold cross validated) with both visual and min SUVr PET metrics. The visual mean AUC of visual and min SUVr were similar (0.81). Combining the two parameters did not improve the overall accuracy (0.79)