Predicting Locoregional Failure After Definitive Radiation Using Machine Learning

Abstract

Clinical staging has been used to predict locoregional recurrence (LR) rates in cohorts of patients undergoing definitive head and neck (HandN) radiation therapy. However, this gross categorization is unable to reliably predict the likelihood of recurrence in an individual patient. High-throughput quantitative imaging features for gross tumor volume derived from pre-treatment multimodality images may provide information that can predict recurrence rates in a given patient, creating a framework towards personalized treatment. We aimed to build a custom machine learning model that can reliably predict the rate of LR prior to the definitive radiation therapy. We retrospectively pooled a large cohort of patients with head and neck cancer treated with definitive radiation from 2008 to 2018 at a single institution. 59 patients were available for analysis (15 patients with LR). Each patient underwent a planning CT scan, an FDG PET scan, and an attenuation corrected (AC) diagnostic CT scan prior to radiotherapy. The gross tumor volume (GTV) was manually drawn on the planning CT. The AC CT scan was diffeomorphically warped to the planning CT using Advanced Normalization Tools (ANTs: http://stnava.github.io/ANTs/), and the transform was applied to the FDG PET scan. 6321 radiomic features (2077 from each of the three scans) from within the GTV mask were extracted using Pyradiomics package. The extracted radiomics features included first order spatial statistics, shape-based volumetrics, and gray level matrix operations on the original image as well as derived images using a variety of spatial filters. Dimensionality reduction was performed using a logistic regression model, reducing the feature dimensions from 6321 to 2754. The remaining features were then fed into a variety of machine learning models to train (N = 35) and validate (N = 24) the predictive model with a balanced number of recurrences in the training and validation sets. The model performance was evaluated using receiver operating characteristic (ROC) curve. We further identified the most important features that may affect prognosis. The machine learning models were able to significantly predict locoregional recurrence in our cohort of 59 patients. The classifier performance of the random forests model revealed an area under the curve (AUC) of 0.81 +/- 0.13. The top feature weights used a combination of features from all three scans, indicating the need for the multi-modal approach of all three scans. We built machine learning models that can be used to predict locoregional recurrence in patients underwent head and neck radiation using pre-treatment PET and CT scans. This model can be applied to better stratify patients based on pre-treatment images towards personalized treatment. We used automated advanced non-linear registrations and neural networks to improve performance from prior models. The approach can be tailored to optimize medical management using data-driven models.