Session Item

Glioblastomas (GBM) are the most common primary brain tumors in adults with a high lethality. Recurrences in this tumor histology are mostly local, despite aggressive treatments combining surgery, chemotherapy, and radiotherapy (RT). Today, RT is prescribed using a “one fits all” concept, i.e., the same dose is prescribed in the whole Planning Target Volume without consideration of local tumor aggressiveness. We made the hypothesis that GBM would benefit from voxel-scale dose painting. We thus investigated whether deep-learning models could predict recurrence sites based on post-operative anatomical MR images. 

All adult patients treated for a histologically proven GBM between 2008 and 2015 in a single institution using a conventional normofractionated RT scheme within STUPP protocol were included. An additional inclusion criterion was the availability of the following sets of images at both baseline (T_Baseline) and recurrence (T_Recurrence) times: a T1-w axial MRI sequence, a T1-w axial MRI sequence with gadolinium injection, and a T2-w axial FLAIR sequence. Primary surgery consisted of either stereotactic biopsy, subtotal resection, or safe wide resection. In all cases, RT delivered a total dose of 60 Gy in 30 fractions, using either 3D conformal RT or intensity-modulated RT. Recurrence was defined as the confirmed appearance of new contrast uptake at recurrence time compared to baseline. A simple task learning strategy based on a modified 3D U-Net architecture was implemented first, with the goal to identify recurrence areas based on the 3 baseline MRI. Sequential and parallel multitask models aiming at segmenting the enhancing part of the tumor at T_Baseline in addition to recurrence areas were considered also. Dice Loss, Binary Cross-Entropy (BCE) Loss, and Dice/BCE Loss were used for optimization. The dataset was randomly split into training (159 patients) and testing (40 patients) sets. Models were trained from scratch using a five-fold cross-validation procedure. Ensembling of the models obtained for the five folds was used for inference on the test set. Networks’ performances were evaluated using the Dice coefficient (DSC) as well as balanced accuracy, sensitivity, and specificity metrics.

Among the 199 patients evaluated, 140 had local recurrence (70%), and 175 (88%) had at least part of the recurring tumor developing from the edema. Table 1 summarizes the results obtained on the test set. The best model obtained a DSC of 0.201. Models exploiting multitasking did not improve performance. The use of a sum of the two losses (Dice + BCE) did not show either a significant contribution whatever the model.

The poor performance highlights the major difficulty of the task, which cannot be solved from anatomical MRI images alone, except perhaps by drastically increasing the amount of data. The addition of functional sequences such as PET or MRI will allow to conclude on the ability to describe tumor aggressiveness in a pre-radiotherapy setting.