Automatic Segmentation and Classification of Glioblastoma and Solitary Brain Metastasis Using a Deep Learning Model...

Area of Science:

• Neuroimaging and Artificial Intelligence
• Oncology and Radiology

Background:

• Distinguishing glioblastoma (GBM) from solitary brain metastasis (SBM) pre-surgically is critical for treatment planning.
• Multiparametric MRI is essential for characterizing brain tumors, but differentiation remains challenging.

Purpose of the Study:

• To develop and validate a three-dimensional (3D) deep learning (DL) model for automated segmentation and classification of GBM and SBM using MRI.
• To compare the performance of the 3D DL model against 2D DL, radiomics, and human radiologists.

Main Methods:

• A cohort of 314 patients with GBM or SBM was analyzed using multiparametric MRI.
• Tumor segmentation was performed using No-new-UNet (nnU-Net).
• 3D DL, 2D DL, and radiomics models were developed for classification; radiologist performance was assessed with and without AI assistance.

Main Results:

• nnU-Net achieved high segmentation accuracy (Dice scores of 0.917 for GBM, 0.915 for SBM).
• The 3D DL model demonstrated superior classification performance (AUC 0.842) compared to 2D DL (0.687), radiomics (0.720), and radiologists (ranging from 0.527 to 0.862).
• Radiologists' diagnostic accuracy significantly improved when using the 3D DL model as a reference.

Conclusions:

• The MRI-based 3D DL model shows significant potential for automated segmentation and classification of GBM and SBM.
• This AI tool can serve as a valuable decision-support system, enhancing radiologists' diagnostic accuracy and potentially improving patient care.
• The model's performance suggests feasibility for clinical translation.