Improved Glioma Grading Using Deep Convolutional Neural Networks.

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Accurate glioma grading is crucial for effective treatment planning.
• Current standard for glioma grading involves invasive tissue sampling.
• Radiomic features show promise but may not capture all image information.

Purpose of the Study:

• To compare the performance of features learned by a convolutional neural network (CNN) against standard radiomic features for glioma grade prediction.
• To evaluate the potential of deep learning in enhancing glioma grading accuracy.

Main Methods:

• A dataset of 237 glioma patients' MR images was analyzed.
• Tumors were segmented after image preprocessing (resampling, registration, skull-stripping).
• CNN-learned features and radiomic features (used with SVM, random forests, gradient boosting) were compared for grade prediction.

Main Results:

• The CNN-based approach achieved an average accuracy of 87% in glioma grade prediction.
• This represents a 23% improvement over the best radiomic feature-based model (gradient boosting, 64% accuracy).
• CNNs demonstrated superior performance in extracting discriminating features for grading.

Conclusions:

• CNNs can automatically learn valuable features for glioma grading, offering added value beyond traditional methods.
• The proposed CNN framework shows potential for substantial improvement in glioma grade prediction.
• Further validation is recommended to confirm the clinical utility of this approach.