Glioma Grading on Conventional MR Images: A Deep Learning Study With Transfer Learning

Area of Science:

• Neuro-oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate preoperative glioma grading is crucial for treatment planning and prognosis.
• Previous magnetic resonance imaging (MRI) studies had limited effectiveness for glioma grading.
• Convolutional neural networks (CNNs) show promise in medical image analysis.

Purpose of the Study:

• To investigate the efficacy of deep learning algorithms, specifically CNNs, for distinguishing World Health Organization (WHO) low-grade from high-grade gliomas using MRI.
• To compare the performance of different CNN architectures (AlexNet, GoogLeNet) and training strategies (from scratch vs. transfer learning).

Main Methods:

• Retrospective analysis of MRI scans from 113 glioma patients.
• Tumor segmentation using a region of interest (ROI) encompassing approximately 80% of the tumor.
• Training and fine-tuning of AlexNet and GoogLeNet models, with and without ImageNet pre-training, evaluated using five-fold cross-validation.

Main Results:

• GoogLeNet trained from scratch achieved validation accuracy of 0.867, test accuracy of 0.909, and test AUC of 0.939.
• Transfer learning and fine-tuning improved performance for both AlexNet and GoogLeNet, with AlexNet showing notable gains.
• GoogLeNet consistently outperformed AlexNet, regardless of the training approach.

Conclusions:

• CNNs, particularly when trained with transfer learning and fine-tuning, significantly enhance preoperative glioma grading accuracy.
• Deep learning approaches surpass traditional machine learning methods relying on hand-crafted features for this task.
• The study demonstrates the potential of advanced CNNs for more precise glioma classification.