Introduction
Glioblastoma (GBM) is a highly aggressive brain cancer with a challenging prognosis. While medical imaging, specifically multi-parametric MRI, provides critical insights into tumor characteristics, survival time (ST) is not solely determined by the tumor's initial state. It is significantly influenced by the chosen clinical intervention, such as the extent of surgical resection.
Traditional AI models for GBM survival prediction typically map MRI features directly to survival estimates, often overlooking the causal impact of treatment. Training separate models for different treatment paths is frequently impractical due to data scarcity for certain procedures. This research introduces a unified framework that incorporates treatment data into the prediction process, allowing for personalized, treatment-specific survival inferences and potential "what-if" analysis for surgical planning.
Methodology
The core of the proposed approach is a treatment-conditioned regression model. The model processes two primary inputs: the patient's status (represented by multi-parametric MRIs and segmentation masks) and a "one-hot" vector encoding the intended treatment type.
The technical innovation lies in the application of Adaptive Instance Normalization (AdaIN). Unlike standard approaches that concatenate clinical data at the final fully connected layers, this model uses the treatment vector to compute scale ($t^s$) and bias ($t^b$) parameters. These parameters are applied to feature maps at every convolutional layer. By injecting treatment information early and consistently, the network learns to extract features that are specifically relevant to the outcome under a given treatment scenario. The backbone is a 3D multi-channel convolutional neural network (CNN) optimized to minimize Mean Absolute Error (MAE).
Experimental Setup
The study utilized the BraTS 2020 (Brain Tumor Segmentation) dataset, comprising 236 patients with survival data ranging from 5 to 1,767 days. Imaging data included four MRI modalities: T1, T1-weighted contrast-enhanced (T1CE), T2, and FLAIR, along with tumor segmentation masks. Treatments were categorized into three groups: Gross Total Resection (GTR, n=119), Subtotal Resection (STR, n=10), and No Treatment (NA, n=107).
Results and Discussion
The integration of treatment information via AdaIN yielded significant improvements in prediction accuracy. As shown in the results table, the AdaIN-based fusion consistently outperformed simple concatenation across different input configurations.
The best-performing configuration, which utilized MRIs, segmentation masks, and AdaIN-based treatment injection, achieved a Mean Absolute Error (MAE) of 98.2 ± 5.6 days. This represents a substantial improvement over baseline models that ignore treatment information (124.4 ± 5.3 days).
Key Insights
- Causal Integration: Survival prediction is enhanced when the model accounts for the interaction between tumor pathology and surgical intervention.
- Unified Modeling: Training a single model on the entire dataset allows the network to learn from all treatment groups simultaneously. This is particularly beneficial for underrepresented groups like Subtotal Resection (STR), where limited samples would otherwise hinder standalone model training.
- Early Feature Conditioning: Injecting treatment information into the convolutional layers (via AdaIN) is more effective than late-stage concatenation, enabling more nuanced feature extraction.
Limitations and Future Work
Despite the performance gains, several areas remain for future development:
- Dataset Size: The BraTS 2020 dataset is relatively small, particularly for specific treatment subgroups. Larger, more diverse cohorts are needed for clinical validation.
- Automation: The current reliance on manual segmentation masks suggests a need for integrating automated deep-learning segmentation to create a seamless end-to-end pipeline.
- Multimodal Data: Future iterations could incorporate tabular Electronic Health Record (EHR) data or radiomics features to further refine predictions.
- Architectural Exploration: The authors plan to investigate more advanced backbones, such as Vision Transformers (ViTs), to potentially enhance feature representation.
Summary updated