MRI-Based End-To-End Pediatric Low-Grade Glioma Segmentation and Classification

Abstract

Purpose: MRI-based radiomics models can predict genetic markers in pediatric low-grade glioma (pLGG). These models usually require tumour segmentation, which is tedious and time consuming if done manually. We propose a deep learning (DL) model to automate tumour segmentation and build an end-to-end radiomics-based pipeline for pLGG classification. Methods: The proposed architecture is a 2-step U-Net based DL network. The first U-Net is trained on downsampled images to locate the tumour. The second U-Net is trained using image patches centred around the located tumour to produce more refined segmentations. The segmented tumour is then fed into a radiomics-based model to predict the genetic marker of the tumour. Results: Our segmentation model achieved a correlation value of over 80% for all volume-related radiomic features and an average Dice score of .795 in test cases. Feeding the auto-segmentation results into a radiomics model resulted in a mean area under the ROC curve (AUC) of .843, with 95% confidence interval (CI) [.78-.906] and .730, with 95% CI [.671-.789] on the test set for 2-class (BRAF V600E mutation BRAF fusion) and 3-class (BRAF V600E mutation BRAF fusion and Other) classification, respectively. This result was comparable to the AUC of .874, 95% CI [.829-.919] and .758, 95% CI [.724-.792] for the radiomics model trained and tested on the manual segmentations in 2-class and 3-class classification scenarios, respectively. Conclusion: The proposed end-to-end pipeline for pLGG segmentation and classification produced results comparable to manual segmentation when it was used for a radiomics-based genetic marker prediction model.

Noninvasive molecular subtyping of pediatric low-grade glioma with self-supervised transfer learning

“…pLGG mutational classification has been previously attempted in a few studies, most with manual segmentation-derived and/or pre-engineered radiomics (35)(36)(37)(38), which are known to fail when applied to the external dataset. Radiomic features have been extracted from MRI images and fitted to classifiers models like XGboost and SVM (17,35,36). One study published in preprint, used neural networks to classify BRAF-mutational status in a single institution, though the algorithm required manual segmentation (16).…”

“…Another barrier to clinical usability is that most algorithms require manual tumor segmentation as input, which is resource-intensive and requires specialized expertise. Few studies have been published investigating pLGG BRAF mutation classification using deep learning (16) and a combination of deep-learning and radiomics (17) but all of them present a single institution study and lack external validation.…”

Novel Imaging Approaches for Glioma Classification in the Era of the World Health Organization 2021 Update: A Scoping Review

“…Most studies analyzed adult populations and high-grade gliomas, with only five studies (8%) analyzing pediatric populations [3][4][5][6][7], ten studies (16%) analyzing low-grade glioma [3,5,6,[8][9][10][11][12][13][14] and only four studies (6%) focusing on diffuse midline glioma [15][16][17][18]. Data on the analyzed patient population are shown in Table 2.…”

“…Overall, 41/62 of the reviewed studies (66%) focused on predicting IDH mutation and 1p/19q codeletion status only, while 33 studies (53%) analyzed other molecular subgroups. These were TERT [9,37,[40][41][42][43][44][45][46], ATRX [8,[47][48][49][50][51], H3K27 [4,[15][16][17][18], MGMT [50,[52][53][54][55], P53 [8,16,51,53], CDKN2A/B [12,30,35,56], EGFR [36], chr7/10 [57] and BRAF alterations [3,[5][6][7]. The reported AUC values range from 0.6 to 0.98 for these predictions with an average of 0.82 to 0.9.…”