Abstract
We developed artificial intelligence models to predict the brain metastasis (BM) treatment response after stereotactic radiosurgery (SRS) using longitudinal magnetic resonance imaging (MRI) data and evaluated prediction accuracy changes according to the number of sequential MRI scans. We included four sequential MRI scans for 194 patients with BM and 369 target lesions for the Developmental dataset. The data were randomly split (8:2 ratio) for training and testing. For external validation, 172 MRI scans from 43 patients with BM and 62 target lesions were additionally enrolled. The maximum axial diameter (Dmax), radiomics, and deep learning (DL) models were generated for comparison. We evaluated the simple convolutional neural network (CNN) model and a gated recurrent unit (Conv-GRU)-based CNN model in the DL arm. The Conv-GRU model performed superior to the simple CNN models. For both datasets, the area under the curve (AUC) was significantly higher for the two-dimensional (2D) Conv-GRU model than for the 3D Conv-GRU, Dmax, and radiomics models. The accuracy of the 2D Conv-GRU model increased with the number of follow-up studies. In conclusion, using longitudinal MRI data, the 2D Conv-GRU model outperformed all other models in predicting the treatment response after SRS of BM.
Introduction
When a patient undergoes stereotactic radiosurgery (SRS) for brain metastasis, clinicians face a recurring diagnostic challenge: the tumor often appears to enlarge on initial follow-up MRI scans. This increase in size can represent either tumor progression (treatment failure) or radiation necrosis (a localized inflammatory response indicating treatment success). Distinguishing between these two is critical, as misdiagnosis leads to either unnecessary surgery or dangerous delays in switching treatments. Current clinical standards rely heavily on manual diameter measurements, which often fail to capture the complex, temporal evolution of a lesion.
Methodology
The researchers compared three distinct approaches to predict treatment response using a pipeline that included serial MRI data, preprocessing (re-sampling and intensity normalization), and BM segmentation.
- Maximum Axial Diameter ($D_{max}$): The traditional clinical baseline.
- Radiomics: Extracting high-dimensional features (texture, shape, intensity) using the PyRadiomics library and an XGBoost classifier.
- Deep Learning (DL): Utilizing a Gated Recurrent Unit (Conv-GRU) architecture. Unlike standard CNNs that analyze single snapshots, the Conv-GRU "remembers" features from previous scans to understand the lesion's trajectory over time.
The DL models utilized a CNN feature extractor based on a ResNet-34 architecture to process image patches from the longitudinal scans.
Key Results and Insights
- The Power of Temporal Dynamics: Model accuracy improved incrementally with every additional follow-up scan. The AUC jumped from ~0.67 with one scan to ~0.87 with four, confirming that the temporal evolution of a tumor contains vital diagnostic signals.
- Top Performer: The 2D Conv-GRU achieved a mean AUC of 0.878 on the developmental set and 0.834 on the temporal test set, significantly outperforming $D_{max}$ (AUC 0.822) and Radiomics (AUC 0.748).
- 2D vs. 3D: Counter-intuitively, the 2D model outperformed the 3D version. This is likely because 3D models are more prone to overfitting on medical datasets of this size, while the 2D approach provided better regularization.
- AI vs. Human-Crafted Features: The DL models consistently beat the Radiomics arm, suggesting that automatically learned features are more predictive of radiation necrosis than traditional mathematical texture formulas.
Conclusion
By "watching the movie" of a tumor's evolution rather than looking at a single snapshot, the 2D Conv-GRU model provides a superior prediction of treatment response after SRS. While the study was limited to a single center and relied largely on radiological follow-up for ground truth, it demonstrates the significant potential of temporal deep learning in neuro-oncology. Future work will focus on multi-center validation and integrating these models directly into clinical PACS workflows.
Summary updated