Model Agnostic Explanations to Identify Radio-Biological Associations of Radiomic Features in MRI Scans of Patients with Brain Metastases

Radiomics, which involves extracting quantitative features from medical images, enhances cancer diagnosis, prog-nosis, and treatment prediction. However, a significant challenge is the interpretability of machine learning models, often seen as “black boxes,” hindering clinical application and regulatory approval. We propose a new workflow building on existing algorithms to develop interpretable models using radiomic features from MRI scans of brain metastases. We used the Yale New Haven Database of 195 patients with brain metastases was used. Radiomic features were extracted using the Py Radiomics library and preprocessed to remove low variance and highly correlated features. Models including linear regression, XGBoost, SVR, and elastic net were developed. Model interpretability was assessed using Local Interpretable Model-Agnostic Explanations (LIME) iterated over 1000 runs. The XGBoost model demonstrated su-perior performance, achieving an RMSE of 0.16, R2 of 0.34, and MAE of 0.12 for necrosis-tumor ratio prediction. LIME identified entropy differences, gray level features, and dependence variance radiomic features from the original filter to influence model predictions the most. The use of LIME on bootstrapped samples for model interpretability aids in clinical application, providing insights into significant radiomic features. Model interpretability and transparency in predictions will pave the way for trustworthy models and enhance acceptance.