Related Experiment Video
Updated: May 7, 2026
07:15
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
6.7K
Radiogenomic explainable AI with neural ordinary differential equation for identifying post-SRS brain metastasis
Jingtong Zhao1, Eugene Vaios1, Zhenyu Yang2
1Deparment of Radiation Oncology, Duke University, Durham, North Carolina, USA.
Medical Physics
|January 29, 2025
Summary
A novel heavy ball neural ordinary differential equation (HBNODE) model accurately differentiates brain metastasis radionecrosis from recurrence post-stereotactic radiosurgery. This explainable AI approach integrates imaging, genomic, and clinical data for improved diagnostic accuracy.
Area of Science:
- Artificial Intelligence
- Radiology
- Oncology
Background:
- Stereotactic radiosurgery (SRS) for brain metastases (BMs) can cause radionecrosis, posing a diagnostic challenge.
- Distinguishing radionecrosis from tumor recurrence non-invasively is difficult with conventional imaging.
- Current machine learning models lack explainability, hindering clinical trust and adoption.
Purpose of the Study:
- To develop a novel neural ordinary differential equation (NODE) model for differentiating BM radionecrosis from recurrence.
- To integrate image-deep features, genomic biomarkers, and clinical parameters into a unified latent feature space.
- To enhance AI explainability by visualizing data sample trajectories within this feature space.
Main Methods:
- A heavy ball NODE (HBNODE) model was designed, treating deep feature extraction as a continuous process governed by a second-order ODE.
- The HBNODE model tracked neural network behavior and visualized data sample trajectories in an Image-Genomic-Clinical (I-G-C) space.
- A decision-making field (F) was reconstructed to analyze feature contributions, identifying key predictive intermediate states.
Main Results:
- The HBNODE model demonstrated superior performance in differentiating radionecrosis from recurrence.
- Achieved an ROC AUC of 0.88 ± 0.04, sensitivity of 0.79 ± 0.02, specificity of 0.86 ± 0.01, and accuracy of 0.84 ± 0.01.
- Outperformed a standard deep neural network (DNN) using only imaging features (ROC AUC 0.71) and a combined model without HBNODE (ROC AUC 0.81).
Conclusions:
- The HBNODE model effectively distinguishes brain metastasis radionecrosis from recurrence, improving explainability in AI.
- The model's performance supports its potential for clinical application in diagnosing post-SRS complications.
- This approach offers a promising framework for explainable AI in various medical imaging and radiogenomic analyses.
