CT texture analysis with machine learning for the prediction of disease site associated features and nodal status for head and neck squamous cell carcinoma

Presentation: C054
Topic: Pharynx / Larynx Cancer
Type: Poster
Date: Thursday, April 19, 2018
Session: 9:00 AM - 7:00 PM
Authors: Xiaoyang Liu1, Eugene Yu1, Behzad Forghani2, Almudena Perez-Lara2, Shao Hui Huang1, John Waldron1, Brian O'Sullivan1, Eric Bartlett1, Mark Levental2, Thomas Ong2, Maryam Bayat2, Reza Forghani2
Institution(s): 1University of Toronto, 2McGill University

PURPOSE: CT plays an essential role in the initial evaluation of head and neck squamous cell carcinoma (HNSCC). We aim to establish a prediction model using quantitative CT texture features to stratify HNSCCs based on their anatomical sites and TNM stages.

MATERIALS & METHODS: A retrospective review of pre-treatment contrast enhanced neck CTs was conducted for 76 oral cavity (OC), 87 oropharyngeal (OP), and 86 laryngeal or hypopharyngeal (LH) cancers. Texture analysis was performed using a commercial software (TexRAD®) by manually delineating a region of interest around the largest diameter of the primary tumor. Random forests (RF) models were constructed using various histogram-based texture features for endpoint prediction with internal cross-validation in addition to the use of separate randomly selected training (70%) and validation (30%) sets. Tumor anatomical site and association with the presence or absence of nodal metastasis and tumor stage was evaluated and sensitivity (Sens), specificity (Spec), positive predictive value (PPV), and negative predictive value (NPV) were determined for endpoint prediction.

RESULTS: The texture parameters of the primary tumor demonstrate good prediction power for the anatomic disease site with sensitivity and specificity of 76% and 90% for LH, 88% and 81% for OP, and 72% and 98% for OC, respectively. They also appeared useful for predicting the presence of associated nodal metastasis, with a PPV of 73% and an NPV of 68%. The features were less robust for distinguishing stage I-II from III-IV tumors.

CONCLUSION: Quantitative CT texture features of primary tumors can be used to predict potentially important endpoints, including association with nodal metastasis. The apparently unique texture features of tumors at different major anatomic sites raises the possibility of unique tumor characteristics and may be used to further refine radiomic models in larger samples.

CLINICAL RELEVANCE STATEMENT: CT texture analysis of primary HNSCC tumors provides important information that may enhance diagnostic tumor evaluation for personalized medicine.