Abstracts

RETURN TO ABSTRACT LISTING


2363. Machine Learning Based on CT Radiomic Features Predicts Residual Tumor in Patients with Head and Neck Cancers Treated with Chemoradiotherapy
Authors * Denotes Presenting Author
  1. Edward Florez *; University of Mississippi Medical Center
  2. Toms Vengaloor Thomas; University of Mississippi Medical Center
  3. Candace Howard; University of Mississippi Medical Center
  4. Seth Lirette; University of Mississippi Medical Center
  5. Ali Fatemi; University of Mississippi Medical Center
Objective:
Surveillance imaging of squamous cell carcinoma of the head and neck (HNSCC) in patients treated with chemoradiotherapy suffers from difficulty in differentiating residual disease from radiation changes and inflammation without a tincture of time. Thus, this study assessed HNSCC cancer patients treated with definitive chemoradiation, to differentiate and predict residual pathological disease from radiation changes and inflammation using radiomic features (RadFs) extracted from standard CT images pre- and post-treatment with the help of machine learning (ML).

Materials and Methods:
A HIPAA-compliant, IRB-approved retrospective post-hoc analysis of HNSCC patients treated with definitive chemoRT at our institution between 2006 and 2015 was performed. Thirty-six patients with residual disease on CT scans of the soft tissue of the neck at a two-month interval—either in primary site, nodal stations, or both—were enrolled. Patients with incomplete follow-up CT images (eg. no post-treatment images), or patients treated with another type of therapy or surgery before the follow-up CT scans, as well as patients whose files contained CT images with severe artifacts (eg. motion and dental hardware) and excessive noise, were excluded. CT images from picture archiving and communication system (PACS), and baseline gross tumor volumes (GTVs) from the treatment planning (CT1) scan were transferred to commercial software. An experienced radiologist performed the contours of HNSCC tumors (primary or nodal lesions) in the two-month follow-up CT scan (CT2) and in the CT portion (CT3) of the three-month follow-up PET/CT scan using segmentation tools. Next, GTVs from baseline and post-therapeutic CT images were exported to an in-house algorithm, where the RadFs were extracted from regions and volumes of interest using different analysis methods. Finally, ML models, including support vector machine (SVM), neural network (NN), and random forest (RF) were used to identify the RadFs, quantifying the changes and progression in HNSCC patients treated with chemoradiotherapy.

Results:
For the 2D scheme from SVM models, RadFs extracted from CT2 had capable predictive ability to anticipate residual disease on PET/CT exams (AUC=0.702). RadFs extracted from PET/CT had moderate predictive ability to predict positive pathology for residual tumor (AUC=0.667). For the 3D scheme using NN and, RF models, RadFs extracted from CT2 and PET/CT had good and moderate predictive ability to predict positive pathology for residual tumor (AUC=0.720 and 0.678, respectively).

Conclusion:
RadFs extracted from both pre- and post-treatment CT data using ML prediction models were able to discriminate between residual tumor and radiation changes in a small group of HNSCC cancer patients treated with chemoradiotherapy. This approach could help to identify patients with a high probability of recurrent disease who may benefit from aggressive early salvage treatment while preventing other patients from receiving unnecessary and invasive treatment.