A Fully Automatic Procedure for Brain Tumor Segmentation from Multi-Spectral MRI Records Using Ensemble Learning and Atlas-Based Data Enhancement

Ágnes Győrfi,László Szilágyi,Levente Kovács

doi:10.3390/app11020564

Ágnes Győrfi, László Szilágyi + Show 1 more

Open Access

https://doi.org/10.3390/app11020564

Copy DOI

Abstract

The accurate and reliable segmentation of gliomas from magnetic resonance image (MRI) data has an important role in diagnosis, intervention planning, and monitoring the tumor’s evolution during and after therapy. Segmentation has serious anatomical obstacles like the great variety of the tumor’s location, size, shape, and appearance and the modified position of normal tissues. Other phenomena like intensity inhomogeneity and the lack of standard intensity scale in MRI data represent further difficulties. This paper proposes a fully automatic brain tumor segmentation procedure that attempts to handle all the above problems. Having its foundations on the MRI data provided by the MICCAI Brain Tumor Segmentation (BraTS) Challenges, the procedure consists of three main phases. The first pre-processing phase prepares the MRI data to be suitable for supervised classification, by attempting to fix missing data, suppressing the intensity inhomogeneity, normalizing the histogram of observed data channels, generating additional morphological, gradient-based, and Gabor-wavelet features, and optionally applying atlas-based data enhancement. The second phase accomplishes the main classification process using ensembles of binary decision trees and provides an initial, intermediary labeling for each pixel of test records. The last phase reevaluates these intermediary labels using a random forest classifier, then deploys a spatial region growing-based structural validation of suspected tumors, thus achieving a high-quality final segmentation result. The accuracy of the procedure is evaluated using the multi-spectral MRI records of the BraTS 2015 and BraTS 2019 training data sets. The procedure achieves high-quality segmentation results, characterized by average Dice similarity scores of up to 86%.

Highlights

Cancers of the brain and the central nervous system cause the death of over two hundred thousand people every year [1]
The goal of this paper is to propose a solution to the brain tumor segmentation problem that produces a high-quality result with a reduced amount of computation, without needing special hardware that might not be available in underdeveloped countries, employing classification via ensemble learning assisted by several image processing tasks designed for Magnetic resonance imaging (MRI) records that may contain focal lesions
To compute the two coefficients of the linear transform applied to the intensities from a certain data channel of any record, first we identify the 25th-percentile and 75th-percentile value of the original intensities and denote them by p25 and p75, respectively

Summary

Introduction

Cancers of the brain and the central nervous system cause the death of over two hundred thousand people every year [1]. Life expectancy after the diagnosis depends on several factors like: being a primary tumor or a metastatic one; being an aggressive form of tumor ( called high-grade glioma (HGG)) or a less aggressive one (low-grade glioma (LGG)); and a key factor is how early the tumor is diagnosed [2]. HGG live fifteen months on average after diagnosis. With an LGG, it is possible to live for several years, as this form of the tumor does not always require aggressive treatment immediately after the diagnosis. Magnetic resonance imaging (MRI) is the technology that has become the most frequently utilized in the diagnosis of gliomas. MRI is preferred because it is much less invasive than other imaging modalities, like positron emission tomography (PET) or computed tomography (CT).

Objectives

Methods

Results

Discussion

Conclusion