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

Purpose/Objective(s): To develop an explainable deep learning (DL) model for accurately distinguishing NSCLC brain metastasis (BM) post-SRS radionecrosis (RN) from true recurrence (TR). By integrating MR image and clinical/genomic data into a deep neural network (DNN), the model dynamically analyzes deep features for diagnosis while forming a risk score, explained by a unique deep space visualization.

Materials/Methods: A Heavy-Ball Neural Ordinary Differential Equation (HBNODE) DL framework, governed by a 2^nd order ODE within a DNN, was designed for this task. The framework enabled dynamic tracking of input evolution within DNN, integrating MR features (extracted by a custom encoder), clinical, and genomic features into a unified Image-Genomic-Clinical (I-G-C) deep feature space. This allowed visualization of data sample trajectories within the space during DNN execution. A novel Layer-Wise Relevance Propagation (LRP) was applied to quantify individual non-imaging feature contributions, capturing their dynamic influence on the final diagnosis.

Within the I-G-C space, a decision-making field (F) was reconstructed, where gradient vectors guided sample trajectories, and potential intensities quantified feature contributions at intermediate states. The temporal evolution of F enabled a quantitative comparison of cumulative contributions from each feature. Key intermediate states, defined as locoregional equilibrium points (?F=0), were identified and aggregated using a non-parametric model to optimize outcome prediction. High-contributing features were selected via k-means clustering of LRP results, forming a risk score model for RN vs. TR differentiation.

Dataset included 142 BM lesions from 103 NSCLC patients, incorporating 3-month post-SRS T1+C MRI, seven genomic biomarkers, and seven clinical parameters. An 8:2 ratio was used for training and independent testing.

Results: The derived risk score integrates 3 high-contributing features via LRP: Age (x1), ALK (x0.84) and PDL-1 (x0.76) status. The Risk Score model outperformed 1) the model using all unweighted non-imaging features and 2) the MR-based DNN model, achieving higher ROC AUC and accuracy with balanced sensitivity and specificity. The HBNODE model, embedding the risk score within deep space, achieved the best performance across all metrics.

Conclusion: The derived risk score, relying solely on non-imaging features, serves as a simple and rapid indicator for differentiating RN from TR. While effective alone, its integration with T1+c MRI in the HBNODE model resulted in the highest predictive performance. The HBNODE model demonstrated superior diagnostic accuracy and explainability, underscoring its potential as a valuable AI-driven tool for BM management.

Table 1

	ROC AUC	Accuracy	Sensitivity	Specificity
‘All C+G’	0.62±0.13	0.69±0.10	0.40±0.15	0.84±0.08
‘MR-only’ DNN	0.71±0.05	0.61±0.11	0.66±0.32	0.59±0.27
Risk Score	0.76±0.03	0.73±0.02	0.71±0.03	0.76±0.04
HBNODE	0.85±0.04	0.82±0.01	0.78±0.01	0.84±0.02