The clinical problem the researchers set out to solve is a timing one. In current practice, the response of a glioblastoma to treatment is assessed with anatomical MRI performed up to three months after the start of radiotherapy and chemotherapy, with lesion sizes measured according to the Response Assessment in Neuro-Oncology, or RANO, criteria. That delay is dangerous in a disease that progresses so quickly, and early interpretation is further complicated by pseudoprogression, an inflammatory phenomenon that can make a responding tumor look as though it is growing. A suspected progression must typically be confirmed by a follow-up scan four to eight weeks later, or by tissue analysis, delaying any definitive judgment. By the time doctors know whether a therapy is working, precious weeks of a patient’s short median survival have already elapsed.

The French team, led by Noémie Moreau and Aurélien Corroyer-Dulmont of the Centre François Baclesse and the University of Caen Normandy, took a different approach: predict the outcome before treatment starts, using only data that already exist in the clinical workflow. They assembled a retrospective cohort of 175 patients treated for glioblastoma between 2018 and 2023, from an initial pool of 385 patients. After excluding clinical trial participants, patients with multiple courses of radiotherapy, those missing pre-treatment MRI scans or clinical data, and patients who were still alive or lost to follow-up, the final cohort comprised 1,575 contrast-enhanced T1-weighted MRI slices from 36,400 acquired images. Importantly, these were routine clinical scans, not the curated, well-organized images from public research datasets that dominate much of the AI literature and may not reflect the messiness of real-world practice.

On the technical side, the researchers fine-tuned a ResNet-51Q convolutional neural network that had been pretrained on ImageNet, a large database of natural images, using transfer learning to adapt it to brain MRI. For each patient, nine central slices through the region of maximum tumor diameter, representing roughly 20 percent of the delineated tumor volume, were selected, and voxel intensities were normalized to compensate for the arbitrary intensity scale of MRI. The network was trained to perform survival classification as a discrete prediction task: first distinguishing short-term from long-term survivors at a threshold of 332 days, the median survival of the training cohort, then tackling three classes defined by terciles, and finally four classes defined by quartiles. Only deceased patients were included, because the classification framework requires exact survival labels that right-censored, still-living patients cannot provide.

To guard against data leakage and inflated performance, the dataset was split at the patient level into training, validation and test sets of roughly 60, 20 and 20 percent, so that no patient’s slices appeared in more than one subset. Hyperparameters were optimized with Optuna, a Bayesian optimization framework, using only the validation set, while the test set was held out until final evaluation. Because the model processes individual slices, the nine per-slice predictions were aggregated into a single patient-level verdict by averaging the probability vectors. Confidence intervals were estimated by patient-level bootstrap resampling with 2,000 iterations, an approach that respects the fact that slices from the same patient are not independent observations. All code was released publicly on GitHub.

The results are striking. In the binary task, the MRI-only deep learning model achieved 81.08 percent accuracy, 94.44 percent specificity and 68.42 percent sensitivity, with an area under the receiver operating characteristic curve of 0.89. The three-class model performed even better at the patient level, reaching 87.18 percent accuracy, 87.18 percent sensitivity and 93.59 percent specificity, with a macro-averaged AUC of 0.948. Even at the more demanding slice level, the three-class model achieved 81.77 percent accuracy. For context, a simple non-AI baseline using surgery type alone, a known prognostic factor, achieved only 69.71 percent accuracy, confirming that the deep network was extracting genuinely predictive information from the images themselves, including features invisible to the human eye such as subtle tissue textures and spatial heterogeneity in vascularization, edema and necrosis.

The team then asked whether adding clinical variables would push performance higher. Using a deep hybrid learning, or DHL, architecture, they concatenated the deep learning model’s averaged probability outputs with structured clinical data: WHO performance status, age, sex and surgery type. This augmented feature vector was fed to a random forest meta-classifier, trained on out-of-fold deep learning predictions generated via five-fold cross-validation to prevent overconfidence bias. The hybrid model reached 84.62 percent accuracy on the three-class task, slightly lower than the 87.18 percent of the MRI-only model, and 76.32 percent on the four-class task versus 71.05 percent for imaging alone. Rigorous statistical testing, including an exact McNemar test and paired bootstrap comparisons, found no significant difference between the two approaches, and feature importance analysis revealed that the imaging contribution dwarfed everything else, accounting for 65 percent of predictive weight in the three-class model and 67.3 percent in the four-class model, with age and surgery type trailing far behind at roughly 13 percent each.

That the imaging signal proved so dominant is scientifically interesting in itself. The authors suggest that surgery type may be partly redundant with imaging, since post-surgical changes such as the resection cavity are visible on the pre-chemoradiotherapy scan and may already be captured by the network. A sensitivity analysis excluding a single patient reclassified under the updated WHO 2021 taxonomy produced only a minor change in accuracy, indicating the result was robust. Still, the team is candid about limitations: the effective independent sample size is 175 patients, the test cohorts for the three- and four-class tasks were small, no external validation on independent multicenter data has yet been performed, and the model was trained on 2D slices rather than full 3D volumes, which could capture more of the tumor’s heterogeneity. The absence of molecular data such as MGMT methylation status, a well-established prognostic marker, also leaves prognostic information unexploited.

Compared with prior work, the Caen study occupies a distinctive niche. Earlier efforts relied on public benchmarks such as the BraTS challenge, used preoperative rather than pre-chemoradiotherapy imaging, or required multimodal data including resting-state functional MRI. One multicenter study predicted survival at eight months after radiotherapy, but used post-treatment MRI, which limits personalization since the therapy has already been given. Others achieved higher reported accuracies, such as 90.6 percent with multimodal preoperative data, but those inputs are not available at the moment of radiotherapy planning, roughly 13 days before treatment begins in the French workflow. By restricting itself to a single contrast-enhanced T1 sequence plus four routine clinical variables, the new model is designed for rapid clinical integration rather than benchmark supremacy.

The researchers emphasize that the tool is not meant to replace the judgment of a radiation oncologist but to provide additional prognostic information at the pre-treatment stage, potentially flagging high-risk patients before any therapy is delivered. Future work includes external validation on multicenter cohorts with harmonized acquisition protocols, federated learning to train more generalizable models without moving sensitive patient data across institutions, deeper multimodal fusion methods, and the integration of radiotherapy dose distributions to capture the effect of the delivered treatment itself. If validated, a system that forecasts survival from the scan already sitting in the patient’s file could transform prognostic stratification for a disease where, for most patients, every week counts.

Subject of Research: Deep learning prediction of overall survival in glioblastoma patients from pre-chemoradiotherapy MRI and clinical data

Article Title: Deep learning prediction of overall survival in glioblastoma patients using pre-chemoradiotherapy MRI and clinical data

Article References: Moreau, N. N., Desmonts, A., Jaudet, C., Leleu, T., Leclercq, A. G., Brunaud, C., Stefan, D., Valable, S., Lechervy, A., & Corroyer-Dulmont, A. (2026). Deep learning prediction of overall survival in glioblastoma patients using pre-chemoradiotherapy MRI and clinical data. Frontiers in Radiology, 6, Article 1963539. https://doi.org/10.3389/fradi.2026.1963539

Image Credits: AI Generated

DOI: 10.3389/fradi.2026.1963539

Keywords: glioblastoma, deep learning, MRI, survival prediction, artificial intelligence, radiotherapy, chemoradiotherapy, deep hybrid learning, neuro-oncology, prognostic stratification, transfer learning, pseudoprogression