1278. PET-CT Radiomic Features Can Predict Clinical Outcomes in Locally Advanced Esophageal Squamous Cell Carcinoma (ESCC)
Authors * Denotes Presenting Author
  1. Vetri Sudar Jayaprakasam *; Memorial Sloan Kettering Cancer Centre
  2. Peter Gibbs; Memorial Sloan Kettering Cancer Centre
  3. Natalie Gangai; Memorial Sloan Kettering Cancer Centre
  4. Ramon Sosa; Memorial Sloan Kettering Cancer Centre
  5. Geoffrey Ku; Memorial Sloan Kettering Cancer Centre
  6. Marc Gollub; Memorial Sloan Kettering Cancer Centre
  7. Viktoriya Paroder; Memorial Sloan Kettering Cancer Centre
The purpose of this study was to assess the usefulness of radiomics features of Positron Emission Tomography – Computed Tomography (PET CT) in patients with locally advanced locally advanced esophageal squamous cell carcinoma (ESCC) in predicting clinical outcomes such as nodal status and progression free and 3-year survival.

Materials and Methods:
This HIPPA compliant study was approved by the institutional review board with waiver for written consent. Patients with locally advanced ESCC without distant metastases who had undergone induction chemotherapy and concurrent chemoradiotherapy between July 2002 and February 2017 were identified. Of the 99 patients identified, twelve patients were excluded due to esophageal stents in situ (3), contrast studies (2), inadequate image quality (5) and non-avid primary tumors (2). Both, pre-treatment PET and CT images were segmented by three radiologists using Hermes Gold LX v2.3.0. Outcome parameters reviewed via the electronic medical record database included progression free survival (PFS), 3-year overall survival and nodal disease. CT images were reduced to 64 grey levels and the PET images were reduced to 16 levels. For the PET data, only those with more than 100 pixels in the region of interest were included. Radiomic features were calculated using Computational Environment for Radiological Research (CERR) [1]. One hundred and one features were calculated in six classes. Adaptive synthetic sampling was employed to equalize class sizes [2]. Least absolute shrinkage and selection operator (LASSO) regression was utilized to determine which radiomic features were of most importance. A maximum of 10 features were selected and predictive models were developed in Matlab using a weighted k nearest neighbor network and 5-fold cross-validation.

A total of 87 PET CTs (58 males, 29 females, mean age 65 years, range 47- 87 years) were segmented. The average inter-reader agreeability using dice score was calculated at 0.776 for the CT and 0.883 for the PET. Both CT and PET data independently showed significant differences in radiomic features in patients with progression compared to no-progression (CT area under the curve [AUC] 0.86, PET AUC 0.80) and node negative compared to node positive patients (CT AUC 0.81, PET AUC 0.92). Combining the CT and PET data, the accuracy for the progression vs no progression group was 78.7% with an AUC of 0.89 and for the node positive vs node negative group, 85.2% with AUC of 0.90. PET demonstrated greater accuracy (76.9%) than CT (57.7%) in predicting 3-year overall survival. Combining PET and CT data further improved accuracy (78.3%).

Radiomics assessment of locally advanced ESCC can help predict clinical outcomes such as progression free survival, nodal status and overall survival. The CT data gives greater accuracy than PET data for the progression outcome whereas the PET data was more accurate for predicting nodal disease and 3-year overall survival. The combined PET and CT data led to slightly improved accuracy for the progression and 3-year overall survival prediction.