Automatic Skull Stripping in Computed Tomographic Images Based on K-Means Statistical Classifier for Radiation Therapy Planning

Abstract

Purpose/Objective(s): PET/CT has a key role in the management of esophageal cancer; however, improved prediction models would be useful for clinical decision making. In particular, ability to predict pathologic complete response (pCR) to chemoradiation would help spare patients from an esophagectomy, a high-risk procedure with significant morbidity. Machine learning is a promising approach to develop models that can predict unseen data, but the efficacy of these models in predicting the onset of esophageal cancer has not been established. We assessed the classification accuracy of prediction models for pathologic complete response based on commonly used machine learning techniques of quantitative pretreatment PET/CT features. Materials/Methods: We identified 39 patients with esophageal cancer who completed neoadjuvant chemoradiation followed by surgery (trimodality therapy) and had pretreatment PET/CT at our institution from 2007-2015. PET/CT images were rigidly fused with radiation planning CT scans based on spine anatomy at the level of gross disease. The following quantitative PET/CT features were evaluated in the clinical target volume (CTV): mean uptake, max uptake, total lesion glycolysis (TLG), and volume. Based on these 4 features, 5 commonly used machine learning techniques (i.e., k nearest neighbors, decision tree, support vector machines, Naive Bayes, decision analysis) were employed to predict the pathologic response (binary classification). Using a leave-one-out cross-validation, the resulting prediction accuracies of these techniques were compared. Results: There were 29 male (74%) and 10 female patients included in this study (median age 61 years, 85% adenocarcinoma, 15% squamous cell carcinoma, all patients stage IIB-IIIC). Median follow-up was 17 months. Pathologic complete response was achieved in 10 patients (26%). Most disease recurrences were distant (13 patients), with 1 local recurrence. Median overall survival was 17 months. Across the 5 machine learning techniques, the average classification accuracy of predicting the pathological response was 0.60 (range, 0.47-0.74) with k nearest neighbors having greatest accuracy at 0.74. Conclusion: Prediction models with fair-to-good prediction accuracy can be developed using simple quantitative pretreatment PET/CT features for patients undergoing trimodality therapy for esophageal cancer. K nearest neighbors may be an especially appropriate method due to robustness to nonlinear data. These data suggest promise in applying machine learning algorithms on quantitative PET data to aid clinical decision making in larger patient populations. Further prospective studies are needed to validate this dataset.