A Robust Semi-Supervised Brain Tumor MRI Classification Network for Data-Constrained Clinical Environments.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Accurate brain tumor classification from MRI is crucial for diagnosis but hindered by labor-intensive manual annotation.
• Developing automated methods is essential for efficient analysis of large medical imaging datasets.

Purpose of the Study:

• To present SSPLNet (Semi-Supervised Pseudo-Labeling Network), a deep learning framework for robust MRI-based brain tumor classification.
• To address data scarcity in clinical settings by leveraging semi-supervised learning and feature fusion.

Main Methods:

• SSPLNet employs a dual-branch deep learning architecture integrating a custom CNN and ResNet50.
• It utilizes confidence-guided iterative pseudo-labeling with adaptive thresholds to refine labels for unlabeled MRI scans.
• Feature fusion combines localized texture patterns and hierarchical deep features for enhanced classification.

Main Results:

• SSPLNet achieves state-of-the-art accuracy across various labeled-unlabeled data splits, outperforming supervised methods in low-label regimes.
• The framework maintains 98.17% diagnostic accuracy with 40% unlabeled data, demonstrating reduced annotation dependence.
• Performance metrics confirm SSPLNet's statistically significant accuracy and robustness, with error rates not attributable to chance.

Conclusions:

• SSPLNet offers a viable solution for scalable brain tumor classification in resource-limited healthcare settings.
• The semi-supervised approach effectively reduces the need for extensive manual data annotation.
• The framework's robust performance and statistical validation support its clinical applicability.