Prediction of treatment response in a longitudinal glioblastoma dataset using deep learning

Motivation: The assessment of treatment response in glioblastoma is a laborious challenge that can be hindered by many confounding phenomena like pseudoprogression on MRI scans. Goal(s): To explore deep learning strategies for classifying treatment response of glioblastoma according to the Response Assessment in NeuroOncology (RANO) criteria using consecutive MRI scans. Approach: We used the longitudinal glioblastoma dataset LUMIERE to compare the performance of 5 different deep learning strategies for the prediction of RANO criteria. Results: The model that yielded the best performance was the DenseNet264, using 3 MRI modalities (T1, T2, and FLAIR) in two consecutive timepoints. Impact: This work sheds light on the deep learning strategies for predicting treatment response in glioblastomas, highlighting the approaches that perform best, thus providing valuable insights into the optimization of the prognostic accuracy of these models.