Predicting Growth Trajectory in Vestibular Schwannoma From Radiomic Data Using Supervised Machine Learning Techniques. (1st September 2019)

• Record Type: Journal Article
• Authors: Grady, Conor; Wang, Hesheng; Schnurman, Zane; Qu, Tanxia; Kondziolka, Douglas
• Is Part Of: Neurosurgery. Volume 66(2010)Supplement 1
• Publication Date: 2019-09-01
• Subjects: Nervous system -- Surgery -- Periodicals; 617.48005
• DOI: 10.1093/neuros/nyz310_318
• Languages: English
• ISSNs: 0148-396X

Abstract

Abstract: INTRODUCTION: Clinicians are not able to predict the growth-rate of a vestibular schwannoma (VS) by reviewing a standard MRI. Recently, the field of radiomics has enabled high-dimensional, quantitative datasets to be created from imaging obtained during routine clinical care. This study investigates whether supervised machine learning techniques can yield accurate predictions of volumetric growth-rate based on radiomic data from MRIs of treatment-naïve VSs. METHODS: A total of 212 patients diagnosed with unilateral VS between 2012 and 2018 underwent measurement of tumor volume on all pre-treatment MRIs. The number of MRIs per patient ranged from 2 to 11, totaling 699 individual studies. Annualized volumetric growth-rate was calculated for each patient. Two cohorts were formed from the 30 patients with the lowest (−20% to 10%) and the 40 patients with the highest (55%–165%) annualized growth-rates, respectively. Manual segmentation of tumor volumes on the last pre-treatment MRI for each patient was performed using 3D Slicer. Pyradiomics was used to calculate histogram, shape, and texture parameters from ADC, CISS, T2 weighted, and postcontrast T1 weighted sequences, resulting in a total of 311 radiomic parameters per volume of interest. Two models predicting cohort membership, a random forest classifier (RFC) and a gradient boosted trees (XGBoost) algorithm, were then trained on a training dataset containing the radiomic profiles of 25 patients from the low growth-rate cohort and 35 patients from the high growth-rate cohort. The models were then tested against the radiomic profiles of the 10 patients withheld from the training group. RESULTS: Following tuning of hyperparameters, both models were able to predict individual tumor membership in the low-growth-rate or high growth-rate cohorts with 100% accuracy. Area under the receiver operating curve (ROC) curve (AUC) was 1.0 for both models. CONCLUSION: Supervised machine learning techniques can predict growth-rate in VS based on radiomic parameters. External validation is warranted.