Differentiation of Low and High-Grade Meningiomas Using Transfer Learning on MR Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Magnetic resonance (MR) imaging aids in diagnosing intracranial tumors like meningiomas, but non-invasive grading remains difficult.
• Accurate meningioma grading is crucial for treatment planning and prognosis.
• Existing methods for tumor grading can be invasive and time-consuming.

Purpose of the Study:

• To explore transfer learning and data augmentation for developing a computer-aided diagnosis (CAD) system for meningioma grading.
• To automatically differentiate between low and high-grade meningiomas using MR images.
• To identify the optimal AI-based solution for meningioma grading.

Main Methods:

• Utilized four MR imaging modalities (T1-weighted, T2-weighted, FLAIR, T1-contrast enhanced).
• Employed a multi-stage framework including image preprocessing and data augmentation (conventional geometric transformations, fancy PCA).
• Trained pre-trained convolutional neural networks (CNNs) using Adam and SGD optimizers.

Main Results:

• The VGG-19 model, trained with fancy PCA-generated images, achieved the highest accuracy (97.80% with SGD, 98.90% with Adam).
• The developed system demonstrated reproducible results on a separate small-scale dataset.
• Fancy PCA data augmentation significantly improved classification performance.

Conclusions:

• Developing CAD systems with pre-trained CNNs and fancy PCA is a promising, non-invasive approach for meningioma grading.
• This AI-based method can assist clinicians in accurate pre-treatment diagnosis.
• The system offers a potential alternative to invasive grading procedures.