CEST Radiomics for Accurate Differentiation of True Tumour Progression from Treatment Effects in Brain Metastases Using Machine Learning

Motivation: Brain metastases are common and difficult to diagnose accurately post-stereotactic radiosurgery (SRS), particularly in distinguishing true tumor progression (TP) from radiation necrosis (RN) non-invasively. Goal(s): To develop a machine learning (ML) model using chemical exchange saturation transfer (CEST) imaging and MRI sequences to differentiate TP from RN. Approach: Radiomic features from APT, NOE, and asymmetry CEST images, along with FLAIR, T1, and T2 MRI sequences, were extracted to train ML models for TP and RN classification. Results: The Extra Trees model with APT and asymmetry-based MTR features achieved high accuracy (AUC=0.86), showing AI's potential to improve non-invasive diagnostics in clinical practice. Impact: These findings suggest that AI-driven analysis of combined CEST and MRI features could serve as an effective diagnostic tool in clinical practice, potentially improving patient management without the need for further invasive procedures.