The BraTS dataset is not divided into Low and High grade gliomas according to the WHO.
We propose a new radiological groundtruth for the BraTS dataset.
Prediction scores give insights about an automatic classifier's behaviour.
The prediction scores histogram can be used as a data exploration tool.
Crossed prediction scores allow comparison of two classifiers and individual tracking.
Glioma grading using maching learning on magnetic resonance data is a growing topic. According to the World Health Organization (WHO), the classification of glioma discriminates between low grade gliomas (LGG), grades I, II; and high grade gliomas (HGG), grades III, IV, leading to major issues in oncology for therapeutic management of patients. A well-known dataset for machine-based grade prediction is the MICCAI Brain Tumor Segmentation (BraTS) dataset. However this dataset is not divided into WHO-defined LGG and HGG, since it combines grades I, II and III as “lower grades gliomas”, while its HGG category only presents grade IV glioblastoma multiform. In this paper we want to train a binary grade classifier and investigate the consistency of the original BraTS labels with radiologic criteria using machine-aided predictions.
Material and methods
Using WHO-based radiomic features, we trained a SVM classifier on the BraTS dataset, and used the prediction score histogram to investigate the behaviour of our classifier on the lower grade population. We also asked 5 expert radiologists to annotate BraTS images between low (as opposed to lower) grade and high grade glioma classes, resulting in a new groundtruth.
Our first training reached 84.1% accuracy. The prediction score histogram allows us to identify the radiologically high grade patients among the original lower grade population of the BraTS dataset. Training another SVM on our new radiologically WHO-aligned groundtruth shows robust performances despite important class imbalance, reaching 82.4% accuracy.
Our results highlight the coherence of radiologic criteria for low grade versus high grade classification under WHO terms. We also show how the histogram of prediction scores and crossed prediction scores can be used as tools for data exploration and performance evaluation. Therefore, we propose to use our radiological groundtruth for future development on binary glioma grading.Le texte complet de cet article est disponible en PDF.
Keywords : Glioma grading, Machine learning, Automatic classification, Prediction score, Virtual biopsy, Radiomics