Predicting response to neoadjuvant chemoradiotherapy in esophageal cancer by textural features derived from pretreatment FDG-PET scans.

Abstract

93 Background: 18-F-fluorodeoxyglucose positron emission tomography/computed tomography (FDG-PET/CT) is currently the imaging method of choice in assessing response of neoadjuvant chemoradiotherapy (nCRT) in esophageal cancer (EC) patients. PET/CT derived texture analysis is potentially more useful than common PET/CT measurements in response assessment and might also be of predictive value in different cancer types. The aim of this study was to develop a model to predict response to nCRT in EC based on pretreatment FDG-PET derived textural features in combination with clinical parameters. Methods: We reviewed 80 locally advanced EC patients who underwent pretreatment FDG-PET/CT and radiation planning CT scans between 2009 and 2015. Patients received nCRT according to the CROSS regimen (carboplatin/paclitaxel/41.4Gy) followed by esophagectomy. We analyzed 7 clinical, 16 geometry-based, and 87 different texture features derived from pretreatment FDG-PET images of the radiotherapy gross tumor volume. Ordinal logistic regression analysis was performed to construct a prediction model for treatment response, which was pathologically classified in complete, partial and no response on the Mandard tumor regression grade (1 vs. 2-3 vs. 4-5). The performance of this model was estimated by comparison with clinical outcome. Results: Pathologic examination revealed 16 (20.0%) complete, 46 (57.5%) partial, and 18 (22.5%) non- responders. Response analysis yielded the following independent predictive textural features: SUVmin, small zone low gray level emphasis, and contrast; and the independent predictive clinical parameters: nodal stage, endoscopic tumor length, and gender. The model has a sensitivity/specificity, positive/negative predictive value, and accuracy of 69%/97%, 85%/93%, and 91% for the prediction of complete response and 61%/79%, 46%/88%, and 75% for non-response, respectively. Conclusions: The prediction model constructed in this study, shows a good overall performance level in predicting response to nCRT in EC patients, but requires further external validation and refinement before it can be used for clinical decision making.