Can Structural MRI Radiomics Predict DIPG Histone H3 Mutation and Patient Overall Survival at Diagnosis Time?

Radiomics was proposed to identify tumor phenotypes noninvasively from quantitative imaging features. The present study aimed at investigating if radiomic features measured at diagnosis time from structural MRI can predict histone H3 mutations and overall survival of patients with diffuse intrinsic pontine glioma. To this end, 316 radiomic features from multimodal diagnostic MRI of 38 patients were extracted, and three clinical parameters were added. Two approaches for computing radiomic features were proposed: a global estimation from a spherical region of interest defined inside the tumor and a local estimation where features are computed inside the previously defined region from fixed size spherical patches and the mean of these features is considered. A feature selection pipeline was then developed. Three machine learning models for H3 mutation classification and three regression models for overall survival prediction were used. Leave-one-out F1-weighted scores for SVM model combining imaging and clinical features reached 0.83, showing a good prediction of H3 mutation using structural MRI. Results on overall survival prediction are not conclusive and suggest the need of a larger number of patients.

I. Introduction

Radiomics [1] is currently widely investigated in oncology. It aims at extracting multiple quantitative imaging features to identify tumor phenotypes with some predictive values. In this study, we investigate the contributions of radiomics to the diagnosis and prognosis of patients with diffuse intrinsic pontine glioma (DIPG). DIPG is a rare inoperable lethal pediatric cancer frequently associated with histone H3 mutations (H3.1K27M and H3.3K27M). These mutations are currently identified from biopsy samples and are associated with patient response to therapy [2]. In this context, we analyzed the ability of radiomic models to distinguish H3 mutation types noninva-sively and to predict patient overall survival (OS). The ultimate goal would be to define whether this could avoid biopsy, or at least replace it when it is not feasible, and guide patient care from diagnosis time. For these prediction tasks, two methods for computing imaging features inside a spherical region of interest included in the tumor were tested, a stringent feature selection procedure was implemented and radiomic signatures were built using different machine-learning methods. Table I Characteristics of 38 DIPG patients included in this study

	H3.1	H3.3	WT/unknown
Patients	8	22	4/4
Age (years)
Boys/girls	3/5	9/13	5/3
OS (days)