Prediction of Radiation Therapy Induced Cardiovascular Toxicity from Pretreatment CT Images in Patients with Thoracic Malignancy via an Optimal Biomarker Approach.

RATIONALE AND

OBJECTIVES

Cardiovascular toxicity is a well-known complication of thoracic radiation therapy (RT), leading to increased morbidity and mortality, but existing techniques to predict cardiovascular toxicity have limitations. Predictive biomarkers of cardiovascular toxicity may help to maximize patient outcomes.

METHODS

The machine learning optimal biomarker (OBM) method was employed to predict development of cardiotoxicity (based on serial echocardiographic measurements of left ventricular ejection fraction and longitudinal strain) from computed tomography (CT) images in patients with thoracic malignancy undergoing RT. Manual segmentations of 10 cardiovascular objects of interest were performed on pre-treatment non-contrast-enhanced CT simulation images in 125 patients with thoracic malignancy (41 who developed cardiotoxicity and 84 who did not after RT). 1078 features describing morphology, image intensity, and texture for each of these objects were extracted and the top 5 features among them that were most uncorrelated and showed the best ability to discriminate between cardiotoxicity/ no cardiotoxicity were determined. The best combination among all possible combinations among these 5 features that yielded the highest accuracy of prediction on a training data set was selected, an SVM classifier was then trained on this combination, and tested for prediction accuracy on an independent data set. Prediction accuracy was quantified for the optimal features derived from each object.

RESULTS

The best feature combination based on 5 CT-based features derived from the left ventricle had the highest testing prediction accuracy of 0.88 among all objects. Prediction accuracies over all objects ranged from 0.76-0.88. Heart, Left Atrium, Aortic Arch, Thoracic Aorta, and Descending Thoracic Aorta showed the next best accuracy of 0.84. Most optimal features were texture properties based on the co-occurrence matrix.

CONCLUSION

It is feasible to predict future cardiotoxicity following RT with high accuracy in individual patients with thoracic malignancy from available pre-treatment CT images via machine learning techniques.