Combined molecular subtyping, grading, and segmentation of glioma using multi-task deep learning

Abstract

Background: Accurate characterization of glioma is crucial for clinical decision making. A delineation of the tumor is also desirable in the initial decision stages but is time-consuming. Previously, deep learning methods have been developed that can either non-invasively predict the genetic or histological features of glioma, or that can automatically delineate the tumor, but not both tasks at the same time. Here, we present our method that can predict the molecular subtype and grade, while simultaneously providing a delineation of the tumor.

Methods: We developed a single multi-task convolutional neural network that uses the full 3D, structural, preoperative MRI scans to predict the IDH mutation status, the 1p/19q co-deletion status, and the grade of a tumor, while simultaneously segmenting the tumor. We trained our method using a patient cohort containing 1508 glioma patients from 16 institutes. We tested our method on an independent dataset of 240 patients from 13 different institutes.

Results: In the independent test set, we achieved an IDH-AUC of 0.90, an 1p/19q co-deletion AUC of 0.85, and a grade AUC of 0.81 (grade II/III/IV). For the tumor delineation, we achieved a mean whole tumor Dice score of 0.84.

Conclusions: We developed a method that non-invasively predicts multiple, clinically relevant features of glioma. Evaluation in an independent dataset shows that the method achieves a high performance and that it generalizes well to the broader clinical population. This first-of-its-kind method opens the door to more generalizable, instead of hyper-specialized, AI methods.

Keywords: deep learning; glioma; multi-task; radiomics; segmentation.

Fig. 1

Overview of our method. Pre- and post-contrast T1w, T2w, and T2w-FLAIR scans are used as an input. The scans are registered to an atlas, bias field corrected, skull stripped, and normalized before being passed through our convolutional neural network. One branch of the network segments the tumor, while at the same time the features are combined to predict the IDH status, 1p/19q status, and grade of the tumor.

Fig. 2

Inclusion flowchart of the train set and test set.

Fig. 3

Receiver operating characteristic (ROC) curves of the genetic and histological features are evaluated on the test set. The crosses indicate the location of the decision threshold for the reported accuracy, sensitivity, and specificity.

Fig. 4

Dice scores, Hausdorff distances, and volumetric similarity coefficients for all patients in the test set.