A Neural Network Approach to Identify the Peritumoral Invasive Areas in Glioblastoma Patients by Using MR Radiomics

Jiun-Lin Yan,Stephen J. Price,Tomasz Matys,Chao Li,Anouk van der Hoorn,Natalie R. Boonzaier

doi:10.1038/s41598-020-66691-6

Abstract

The challenge in the treatment of glioblastoma is the failure to identify the cancer invasive area outside the contrast-enhancing tumour which leads to the high local progression rate. Our study aims to identify its progression from the preoperative MR radiomics. 57 newly diagnosed cerebral glioblastoma patients were included. All patients received 5-aminolevulinic acid (5-ALA) fluorescence guidance surgery and postoperative temozolomide concomitant chemoradiotherapy. Preoperative 3 T MRI data including structure MR, perfusion MR, and DTI were obtained. Voxel-based radiomics features extracted from 37 patients were used in the convolutional neural network to train and as internal validation. Another 20 patients of the cohort were tested blindly as external validation. Our results showed that the peritumoural progression areas had higher signal intensity in FLAIR (p = 0.02), rCBV (p = 0.038), and T1C (p = 0.0004), and lower intensity in ADC (p = 0.029) and DTI-p (p = 0.001) compared to non-progression area. The identification of the peritumoural progression area was done by using a supervised convolutional neural network. There was an overall accuracy of 92.6% in the training set and 78.5% in the validation set. Multimodal MR radiomics can demonstrate distinct characteristics in areas of potential progression on preoperative MRI.

Highlights

The imaging characteristics to this peritumoural non-enhanced invasive area is still not clear
Another study focused on the Diffusion tensor imaging (DTI) fractional anisotropy (FA) in the peritumoural non-enhancing area showed that a decrease of FA may indicate tumour infiltration which further lead to local tumour recurrence[10]
Further study conducted by Kim et al, showed that the FA and the CBV radiomics of the peritumoural area, together with the clinical index, can potentially predict the tumour progression[13]

Summary

Introduction

The imaging characteristics to this peritumoural non-enhanced invasive area is still not clear. Another study focused on the DTI fractional anisotropy (FA) in the peritumoural non-enhancing area showed that a decrease of FA may indicate tumour infiltration which further lead to local tumour recurrence[10]. In the conventional imaging analysis, only limited semantic features can be extracted, by using high-throughput computing, radiomics can quantify more detailed agnostic features including first order, second order or higher order. In this study we aimed to characterize the preoperative peritumoural non-enhanced area that demonstrated to have generated tumour progression later on. Further applied these features, together with radiomics features to identify areas of GBM progression on the preoperative MRI by using the convolutional neural network

Objectives

Methods

Results

Discussion

Conclusion