Handcrafted and Deep Learning-Based Radiomic Models Can Distinguish GBM from Brain Metastasis

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Accurate preoperative classification of brain tumors is crucial for treatment planning.
• Distinguishing glioblastoma (GBM) from solitary brain metastasis (BM) can be challenging using conventional imaging alone.

Purpose of the Study:

• To evaluate the feasibility of handcrafted radiomics (HCR) and deep learning-based radiomics (DLR) for preoperative classification of GBM and BM.
• To compare the performance of HCR and DLR models across different MRI sequences.

Main Methods:

• Retrospective analysis of MRI data from 140 GBM and 128 BM patients.
• Manual region of interest (ROI) delineation on T1WI, T2WI, and T1CE MRI sequences.
• Application of HCR and DLR feature extraction, followed by machine learning model implementation and validation.

Main Results:

• Random forest models demonstrated strong performance across MRI modalities.
• HCR models achieved good discrimination (e.g., T1CE AUC=0.93).
• Integration of DLR features significantly improved AUC values for all sequences (e.g., T1CE AUC=0.97).
• The T1CE-based radiomic model, particularly with combined HCR and DLR, achieved the highest classification performance (AUC=0.97 in test set).

Conclusions:

• Machine learning models incorporating MRI radiomic features are effective for GBM vs. BM differentiation.
• The combination of HCR and DLR features offers superior classification performance.
• Radiomics analysis holds significant potential for improving preoperative brain tumor classification.