An active learning approach to train a deep learning algorithm for tumor segmentation from brain MR images.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Neuro-oncology research

Background:

• Brain glioma segmentation from MRI is crucial for diagnosis and treatment planning.
• Deep learning models require extensive annotated data, posing a significant bottleneck.
• Active learning offers a potential solution to optimize data annotation efficiency.

Purpose of the Study:

• To evaluate the effectiveness of active learning techniques in training a brain MRI glioma segmentation model.
• To determine if active learning can reduce the manual annotation effort while maintaining model performance.
• To assess the viability of active learning for improving the efficiency of ground truth data preparation.

Main Methods:

• Utilized the 2021 BraTS Challenge dataset (1251 multi-parametric MRIs).
• Trained deep convolutional neural network segmentation models using the NiftyNet platform.
• Implemented an active learning strategy by iteratively adding model-predicted segmentations to the training set.

Main Results:

• Active learning achieved comparable glioma segmentation performance (Dice score 0.868) to a reference model (Dice score 0.906).
• The active learning approach required manual annotation for only 28.6% of the data.
• Demonstrated a significant reduction in manual annotation effort.

Conclusions:

• Active learning is a viable strategy for training brain MRI glioma segmentation models.
• This method substantially decreases the time and labor associated with ground truth data preparation.
• Active learning enhances the efficiency of developing AI tools for neuro-oncology.