A Transfer Learning-Based Active Learning Framework for Brain Tumor Classification.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Oncology
• Machine Learning for Healthcare

Background:

• Brain tumors are a major cause of cancer deaths globally in children and adults.
• Accurate grading of gliomas (low-grade vs. high-grade) is crucial for prognosis and treatment.
• Deep learning models for brain tumor classification require large annotated datasets, which are difficult to obtain.

Purpose of the Study:

• To develop a novel transfer learning-based active learning framework for brain tumor classification.
• To reduce the cost and effort of data annotation while maintaining model performance.
• To improve the stability and robustness of AI-driven brain tumor grading systems.

Main Methods:

• A 2D slice-based approach was used to train and fine-tune a model on magnetic resonance imaging (MRI) data.
• A retrospective dataset of 203 patients for training and 66 for validation was utilized.
• A transfer learning-based active learning framework was proposed to optimize annotation efficiency.

Main Results:

• The proposed model achieved an Area Under the ROC Curve (AUC) of 82.89% on a test set, outperforming the baseline by 2.92%.
• The method demonstrated a reduction in labeling costs by at least 40%.
• On a balanced dataset, the model achieved an AUC of 82%, compared to 78.48% for the baseline, confirming robustness.

Conclusions:

• The novel transfer learning augmented with active learning framework effectively classifies brain tumors.
• This approach significantly reduces annotation requirements and costs for deep learning models.
• The method offers a robust and stable solution for AI-assisted brain tumor grading, improving diagnostic efficiency.