Deep semi-supervised learning for brain tumor classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Glioma classification from multi-modal MRI is crucial but challenged by unlabeled and moderately sized datasets.
• Existing methods struggle with data scarcity and the effective utilization of unlabeled brain scans.

Purpose of the Study:

• To develop a deep semi-supervised learning framework for enhanced glioma classification.
• To leverage unlabeled MRI data and address overfitting in moderate-sized datasets for improved diagnostic accuracy.

Main Methods:

• Incorporated deep Convolutional Neural Network (CNN) features into a graph-based semi-supervised learning framework.
• Introduced a 3D-2D consistent constraint for slice-level classification consistency within 3D brain scans.
• Utilized Generative Adversarial Networks (GANs) to create synthetic MRIs for augmenting training data and mitigating overfitting.

Main Results:

• Achieved high performance in glioma IDH-mutation prediction (86.53% accuracy on TCGA dataset).
• Demonstrated strong performance in glioma grading (90.70% accuracy on MICCAI dataset).
• The proposed method effectively utilizes unlabeled data and synthetic data augmentation.

Conclusions:

• The deep semi-supervised learning scheme is effective for glioma IDH-mutation prediction and grading.
• The approach achieves performance comparable to state-of-the-art methods in brain tumor classification.
• This framework offers a promising solution for improving glioma classification accuracy with limited labeled data.