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

Area of Science:

• Neuro-oncology
• Medical imaging
• Artificial intelligence

Background:

• Improved survival in brain metastases patients necessitates accurate differentiation between local recurrence (LR) and radionecrosis (RN).
• Stereotactic radiosurgery (SRS) for non-small cell lung cancer (NSCLC) brain metastases presents a challenge in distinguishing LR from RN post-treatment.

Purpose of the Study:

• Develop an explainable deep learning (DL) model for non-invasive differentiation of RN from LR in NSCLC patients post-SRS.
• Integrate multi-modal data including MRI, clinical, and genomic features for enhanced diagnostic accuracy.

Main Methods:

• Designed a Heavy-Ball Neural Ordinary Differential Equation (HBNODE) DL framework for dynamic input evolution tracking.
• Integrated MR, clinical, and genomic features into a unified Image-Genomic-Clinical (I-G-C) space.
• Applied Layer-Wise Relevance Propagation (LRP) to quantify feature contributions and identified key features (age, ALK, PD-L1 status) for a risk score model.

Main Results:

• A risk score model based on age, ALK, and PD-L1 status demonstrated superior performance compared to models using unweighted features or MRI alone.
• The HBNODE model, incorporating the risk score, achieved optimal performance across all evaluated metrics.
• LRP analysis quantified the influence of individual non-imaging features on the diagnostic outcome.

Conclusions:

• A risk score derived from non-imaging features provides a rapid and simple method for distinguishing RN from LR.
• Integration of this risk score with MRI within the HBNODE model improved predictive performance and maintained explainability.
• The developed model shows potential as a clinical decision-aid tool for managing brain metastases.