18F-FDOPA PET for the Noninvasive Prediction of Glioma Molecular Parameters: A Radiomics Study

Abstract

The assessment of gliomas by ¹⁸F-FDOPA PET imaging as an adjunct to MRI showed high performance by combining static and dynamic features to noninvasively predict the isocitrate dehydrogenase (IDH) mutations and the 1p/19q codeletion, which the World Health Organization classified as significant parameters in 2016. The current study evaluated whether other ¹⁸F-FDOPA PET radiomics features further improve performance and the contributions of each of these features to performance. Methods: Our study included 72 retrospectively selected, newly diagnosed glioma patients with ¹⁸F-FDOPA PET dynamic acquisitions. A set of 114 features, including conventional static features and dynamic features, as well as other radiomics features, were extracted and machine-learning models trained to predict IDH mutations and the 1p/19q codeletion. Models were based on a machine-learning algorithm built from stable, relevant, and uncorrelated features selected by hierarchic clustering followed by a bootstrapped feature selection process. Models were assessed by comparing area under the curve using a nested cross-validation approach. Feature importance was assessed using Shapley additive explanations values. Results: The best models were able to predict IDH mutations (logistic regression with L2 regularization) and the 1p/19q codeletion (support vector machine with radial basis function kernel) with an area under the curve of 0.831 (95% CI, 0.790-0.873) and 0.724 (95% CI, 0.669-0.782), respectively. For the prediction of IDH mutations, dynamic features were the most important features in the model (time to peak, 35.5%). In contrast, other radiomics features were the most useful for predicting the 1p/19q codeletion (up to 14.5% of importance for the small-zone low-gray-level emphasis). Conclusion:¹⁸F-FDOPA PET is an effective tool for the noninvasive prediction of glioma molecular parameters using a full set of amino-acid PET radiomics features. The contribution of each feature set shows the importance of systematically integrating dynamic acquisition for prediction of the IDH mutations as well as developing the use of radiomics features in routine practice for prediction of the 1p/19q codeletion.

Keywords: 18F-FDOPA PET; WHO 2016 classification; glioma; machine learning; radiomics.

Graphical abstract

FIGURE 1.

Flowchart of retrospective selection of study patients.

FIGURE 2.

Detailed workflow of image acquisition and reconstruction, VOI segmentation (shown in red contours), and features extracted, where images and VOIs are listed in orange and gray rectangles, respectively, processing steps in green rectangles, intermediate results in dark blue rectangles, and final features in light blue rectangles. GLCM = gray level co-occurrence matrix; GLRLM = gray level run length matrix; GLSZM = gray level size zone matrix; MTV = metabolic tumor volume; NGTDM = neighborhood gray tone difference matrix; NGLDM = neighboring gray level dependence; p.i. = after injection; TBR = tumor–to–normal-brain ratio; TSR = tumor-to-striatum ratio; TAC = time–activity curve; TTP = time to peak.

FIGURE 3.

Modeling pipeline using nested cross-validation to obtain unbiased estimate of model performance. Fifty repeats of 3-fold cross-validation were used as inner loop and 5 repeats of 10-fold cross-validation as outer loop. For each fold of outer cross-validation, hyperparameters of model were tuned in inner cross-validation using only outer training data. Best hyperparameters chosen were then used to fit model to outer training data, and model performance was evaluated on outer test data. Blue and red rectangles, respectively, denote training and test data for each cross-validation. Green arrows represent predictions made by fitted model on test data, and green boxes represent score calculated from these predictions. CV = cross-validation; OF = outer fold; IF = inner fold; Si = score on test data of outer fold i; Si.j = score on test data of inner fold j from outer fold i.

FIGURE 4.

Feature importance derived from outer models for most contributive features to prediction of IDH mutations using logistic regression with L2 regularization and 5 features selected (A) and 1p/19q codeletion using support vector machine with radial basis function kernel and 15 features selected (B). Individual importance of each feature was normalized to sum of 1.0. Other radiomic features are in blue, and dynamic features are in orange.

FIGURE 5.

Overview of impact on model of each feature according to its values (for prediction of IDH mutations [A] and for prediction of 1p/19q codeletion [B]). Impact on model output is shown with SHAP values on x-axis, and feature value is displayed in colors (blue and red for, respectively, low and high values). For instance, in A, for classification of IDH mutations, high values of time to peak (in red) are showing positive SHAP values and thus are associated with prediction of positive class (IDH-positive gliomas) as opposed to low values of time to peak (in blue), which exhibit negative SHAP values, meaning that they are associated with prediction of negative class (IDH-negative gliomas).

FIGURE 6.

Representative examples of patients with IDH wild-type glioma (A) and 1p/19q codeleted glioma (B). Shown for each patient are axial slice of ¹⁸F-FDOPA PET (left), dynamic mean tumor–to–normal-brain ratio curve (middle; time to peak, light gray dotted line; slope, dark blue dotted line), and same slice on FLAIR MRI (right), above graphic representing contribution of features involved in model prediction (IDH mutations classification [A]; 1p/19q codeletion classification [B]). Red features push prediction toward positive class, whereas blue features push toward negative class; longer arrow indicates more impact on model. Base model values (value of featureless model) and final decision values are displayed, respectively, in italics and boldface. For interpretation purposes, feature names along with their value (expressed in z score) are displayed under each arrow. TTP = time to peak; ZP = zone percentage; DV = difference variance; LZE = large-zone emphasis; SZLGE = small-zone low-gray-level emphasis; GLNU = gray-level nonuniformity; DCNUN = dependence count nonuniformity normalized. IS = intensity statistics; GLCM = gray level co-occurrence matrix; GLRLM = gray level run length matrix; GLSZM = gray level size zone matrix; MORPH = morphologic; NGTDM = neighborhood gray tone difference matrix; NGLDM = neighboring gray level dependence matrix.