A one-stage multi-task network for molecular subtyping, grading, and segmentation of glioma

Abstract

The World Health Organization tumor classification emphasizes the key role of molecular biomarkers in glioma diagnosis, particularly the importance of isocitrate dehydrogenase (IDH) mutation status and 1p/19q co-deletion status. There’s little research that combines glioma segmentation with the prediction of their genetic or histological characteristics using multimodal magnetic resonance imaging (MRI) scans. We proposed a one-stage multi-task network that uses MRI scans to predict IDH mutation status, 1p/19q co-deletion status, and glioma grading while simultaneously segmenting tumors. The network features an encoder-decoder architecture with three main components: an encoder that extracts multi-scale features, a decoder that gradually aggregates these features for segmentation, and a masked multi-scale fusion module that merges the features with the segmentation output to perform classification. A multi-task learning loss is then used to balance all tasks. The proposed method was evaluated using a public dataset and a local hospital’s dataset. The results demonstrate that the proposed method achieves superior performance while consuming fewer computational resources compared to existing networks. In the testset of the public dataset, it achieves Area Under Curves (AUC) of 0.9851 (IDH), 0.7695 (1p/19q), and 0.8949 (grade) with a mean dice score of 0.8485 and a mean Hausdorff distance of 19.60 mm; in the local hospital’s dataset, the AUCs were 0.9313, 0.8254, and 0.8638, with a mean dice score of 0.7490 and a mean Hausdorff distance of 24.50 mm. The proposed method can be potentially used in clinical practice to alleviate patient suffering, serving as a diagnostic tool for glioma patients.