P07.17 Automated grade prediction of glioma patients based on magnetic resonance imaging and a random forests approach

Abstract

Introduction: The accurate characterisation of a brain tumour is of primary importance for determining the optimal therapy. Up to now, the diagnosis is determined based on a biopsy, requiring invasive surgery. Furthermore, a certain degree of subjectivity is involved in this approach. Therefore, in recent years several studies have investigated automated tumour classification based on medical imaging. For this purpose, quantitative parameters are extracted from the images describing the tumour appearance. These features are then used as input for a classification algorithm. In this study the goal was to implement an easy classification method based on random forests to distinguish between low-grade and high-grade glioma on conventional MRI scans.Materials and methods: Publicly available brain tumour scans were obtained from the MICCAI BRATS website. This dataset consists of 54 low-grade glioma patients (astrocytoma or oligoastrocytoma) and 220 high-grade glioma patients (anaplastic astrocytoma or glioblastoma multiforme). For every patient, T1, T2, contrast enhanced T1 and FLAIR scans are available, as well as a manual tumour segmentation. Using this segmentation, we extracted 208 quantitative features describing the tumour intensity, shape and texture. These features were next used to train a Random Forests classification algorithm using the Matlab TreeBagger function. For every patients 1000 classification trees were grown and validated using a leave-one-out scheme. However, due to the unbalanced dataset (ca. 4 times more high-grade than low-grade samples), the algorithm will favour high-grade classification. Therefore, we applied two correction techniques: penalising low-grade samples misclassified as high-grade, and Balanced Random Forests, where every tree is grown with an equal number of low- and high-grade samples.Results: With the misclassification cost, we obtained a high-grade sensitivity of 93.2% and specificity of 64.8%, with a global accuracy of 87.6%. Using the Balanced Random Forests method, a high-grade sensitivity of 95.5% and specificity of 79.6%, with a global accuracy of 92.3% was achieved. These results not only indicate that we achieve a similar accuracy as in literature, but that for classification, a relatively simple approach can already give excellent results. Moreover, the performance of this simple automated classification algorithm is significantly higher than the performance of expert readers who had to do the same task on conventional MR imaging. Law et al. (Am J Neuroradiol, 2003) mention an accuracy of 70.6% for classifying a glioma patient as low- or high-grade. Moreover, we are convinced that the inclusion of dynamic MRI or PET features is promising to further improve the classification accuracy.Conclusions: Using an easy classification algorithm, we were able to distinguish low-grade from high-grade glioma based on medical imaging.