An Explainable Deep Model for Risk Scoring and Accurate Radionecrosis Identification Following Brain Metastasis Stereotactic Radiosurgery.

[PURPOSE] As survival improves for patients with brain metastases (BM), distinguishing local recurrence (LR) from radionecrosis (RN) is a growing neuro-oncologic challenge. We aimed to develop an explainable deep learning model to noninvasively distinguish RN from LR in patients with non-small cell lung cancer following stereotactic radiosurgery.

[METHODS AND MATERIALS] A second-order heavy-ball neural ordinary differential equation (HBNODE) deep learning framework was designed. It enabled dynamic tracking of input evolution within a deep neural network, integrating magnetic resonance (MR), clinical, and genomic features into a unified Image-Genomic-Clinical space. This allowed visualization of sample trajectories during model execution. Layer-wise relevance propagation (LRP) was applied to quantify individual nonimaging feature contributions and their influence on diagnosis. Within the Image-Genomic-Clinical space, a decision-making field (F) was reconstructed. The temporal evolution of F enabled quantitative comparison of cumulative contributions from each feature. Key intermediate states, defined as locoregional equilibrium points (∇F = 0), were identified and aggregated using a nonparametric model to optimize prediction. High-contributing features were selected via k-means clustering of LRP results, forming a risk score model for RN versus LR differentiation. The data set included 142 BM lesions from 103 non-small cell lung cancer patients, incorporating 3-month post-stereotactic radiosurgery T1 + C magnetic resonance imaging, 7 genomic biomarkers, and 7 clinical parameters. An 8:2 ratio was used for training and independent testing.

[RESULTS] Three high-contributing features, age (×1), ALK rearrangement status (×0.84), and PD-L1 expression status (×0.76), were identified by LRP and used to constructs the risk score. The risk score model outperformed the model using all unweighted clinical/genomic features and an MR-only deep neural network. The HBNODE model, embedding the risk score within deep space, achieved the best performance across all metrics.

[CONCLUSIONS] The derived risk score, based on nonimaging features, is a simple and rapid indicator for distinguishing RN from LR. When integrated with magnetic resonance imaging in the HBNODE model, it further enhanced predictive performance while maintaining high explainability, highlighting its potential as a clinical decision-aid tool for BM management.