From Radiomics to Generative Models: Evaluating Early Radiation Effects in Metastatic Brain Lesions.

Area of Science:

• Neuro-oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Brain metastases (BM) are common malignant brain tumors requiring advanced imaging for monitoring.
• Stereotactic radiosurgery (SRS) is a key treatment, but radiation effects can complicate diagnosis.
• Early radiation effects (ERE) and adverse radiation effects (ARE) present diagnostic challenges, impacting patient management.

Purpose of the Study:

• To compare standard radiomics machine learning with pretrained generative models for assessing ERE in BM lesions.
• To evaluate the efficacy of multiparametric PET/MRI analysis, radiomics, and generative models in differentiating treatment effects from tumor progression.
• To determine if generative models can improve diagnostic accuracy and reduce the need for manual segmentation.

Main Methods:

• Analysis of multiparametric 18F-FPIA PET/MRI data from 21 patients with BM.
• Comparison of PET/MRI parameters, radiomics (MRI, PET, combined), dynomics, and generative model embeddings.
• Application of machine learning classifiers (SVM, XGBoost, Linear regressor) with cross-validation.

Main Results:

• Radiomic approaches showed comparable performance in assessing ERE (Accuracy: 71.95%, AUC: 0.72).
• Pretrained generative models achieved significantly higher performance (Accuracy: 83.82%, AUC: 0.83).
• Generative models provided superior results without requiring manual lesion segmentation.

Conclusions:

• Generative models offer a more accurate and efficient method for assessing ERE in BM lesions.
• This AI-driven approach has the potential to streamline neuro-oncology imaging analysis.
• The study highlights the clinical relevance of advanced AI in improving diagnostic capabilities for radiation-induced effects.