SVM based Radiomics Analysis using Pre-radiotherapy PET/CT Increases the Prediction Accuracy of Radiation Pneumonitis

Objectives: Radiation pneumonitis (RP) is a severe and potentially fatal complication of thoracic radiotherapy (RT) for breast, esophageal and lung cancer patients. In this study, radiomics features derived from pre-RT FDG-PET/CT images were studied as potential prognostic biomarkers of symptomatic radiation pneumonitis in lung cancer patients. We hypothesize that these image derived features can identify the individuals at high risk for the development of RP and assist the personalized treatment planning. Support Vector Machine (SVM), a novel machine learning tool, was trained using multiple radiomics features and applied to improve the prognosis.

Methods: Forty patients (age 64.8y±8.7, F=14, M=26) with non-small cell lung cancer, who were treated with either intensity modulated RT (N=16) or proton therapy (N=24), were included in this study. Both pre-treatment and post-treatment PET/CT images were acquired for each patient. The tumor region of interest (ROI) was defined by gross tumor volume (GTV) from planning CT images and was transferred to registered PET/CT images. Ipsilateral and contralateral lung ROIs were manually contoured for exclusion of central airway and PET spill-over artifacts as well as cold-spot artifacts near diaphragm. In radiomics analysis, we tested several series of texture and heterogeneity features derived from both PET and CT images. The textures were extracted using Laplacian of Gaussian (LoG) Filter and Gray-Level based matrices (e.g. Co-Occurrence Matrix) on 2D image slices while heterogeneity was characterized by Global Moran I(d) analysis on 3D image volume. We also included the performance of conventional biomarkers respective to SUV_mean and SUV₉₅ defined as the SUV reach the 95% cumulative histogram distribution in the ROI. The total number of radiomics features reached 115 in the evaluation (as show in Table 1). For each image feature, the quantitation derived from PET/CT images acquired before treatment was tested as individual biomarker in differentiating the RP patient group (N=18) from the non-RP group (N=22). The Wilcoxon Rank-Sum test was performed to compare the performance of each feature and statistical significance at p-value < 0.05 was used for feature selection. Further, receiver operating characteristic (ROC) was used to characterize the prediction accuracy of each image feature. Thus the features which can successfully differentiate the two patient groups were identified. Then we trained a SVM to fuse the selected radiomics features for prospective prediction. The trained SVM structure was applied to test the SVM performance in classifying RP/non-RP patient groups using cross-one-validation.

Results: Statistically significant differences between RP/nonRP groups were observed in the grey level mean of pre-treatment PET image texture of ipsilateral lung (p<0.05) as well as in the SUV_mean (p<0.05) and SUV₉₅ (p<0.05) of both ipsilateral and contralateral lungs. The area under the curve in ROC analysis for each individual feature was between 68% to 76% and the prediction accuracy of symptomatic radiation pneumonitis was limited using single feature. When all selected features were used as input of SVM, the accuracy of SVM classification is 85%, significantly better than the SVM accuracies (65%~72.5%) when single feature was used as input.

Conclusion: Radiomics features derived from pre-treatment PET/CT could be used as predictive biomarkers for symptomatic RP. SVM can significantly improve the accuracy of RP prediction with multiple features and provides a more accurate prognostic biomarker to assist the patient specific treatment planning. Research Support: R01EB013293 $$graphic_7C32DCEF-8F6C-4A68-B025-5242B52D548B$$

• Download figure
• Open in new tab
• Download powerpoint