Author: Evan Calabrese, Scott R. Floyd, Kyle J. Lafata, Zachary J. Reitman, Eugene Vaios, Chunhao Wang, Lana Wang, Deshan Yang, Zhenyu Yang, Jingtong Zhao 👨🔬
Affiliation: Duke University, Department of Radiation Oncology, Duke University, Duke Kunshan University 🌍
Purpose:
This study proposes a novel neural ordinary differential equation (NODE) framework to distinguish post-SRS radionecrosis from recurrence in brain metastases (BMs). By integrating imaging features, genomic biomarkers, and clinical parameters into a unified deep feature space, the model enables visualization of deep space evolution, providing a solution for model explainability.
Methods:
A heavy-ball NODE (HBNODE) framework was created to model spatiotemporal continuity in deep feature space evolution using a second-order ODE. This approach enabled visualization of sample trajectories in the Image-Genomic-Clinical (I-G-C) deep feature space and reconstruction of a decision-making field (F), where gradient vectors directed sample trajectories and potential intensities quantify feature contributions at intermediate states. Temporal evolution of F allowed comparison of dynamic contributions from imaging, genomic, and clinical features. Key intermediate states (when ∇F=0) were identified and aggregated using a non-parametric model to predict outcomes. Layer-wise Relevance Propagation (LRP) was adopted to attribute relative contributions of clinical and genomic features, while K-means clustering of the LRP results identified features that contributed the most to the final predictions.
The dataset included 142 BMs from 103 NSCLC patients, with 3-month post-SRS T1+c MR image features, seven genomic biomarkers, and seven clinical parameters. An 8:2 train/test split was used, and models were trained using 5-fold cross-validation to ensure robustness. Sensitivity, specificity, accuracy, and ROCAUC were evaluated.
Results:
The model achieved an ROCAUC of 0.85±0.04, sensitivity of 0.78±0.01, specificity of 0.84±0.02, and accuracy of 0.83±0.01. Feature contributions in deep space evolved, with clinical features dominating early stages and imaging features becoming predominant in later stages. Genomic features contributed comparatively less overall. LRP identified four key high-contributing clinical and genomic features.
Conclusion:
The HBNODE model demonstrated robust performance and explainability, offering a promising AI-driven solution for distinguishing post-SRS radionecrosis from recurrence and advancing explainable AI in clinical practice.