Two Stream Active Query Suggestion for Active Learning in Connectomics.

Area of Science:

• Biomedical image analysis
• Computer vision
• Machine learning

Background:

• Deep learning models require large labeled datasets for optimal performance in large-scale vision tasks.
• Active learning strategies are employed to mitigate limited labeled data by iteratively selecting informative unlabeled samples for annotation.
• Existing query suggestion methods in active learning often underperform on challenging unlabeled data due to optimization solely on limited labeled sets.

Purpose of the Study:

• To develop an improved query suggestion approach for active learning in biomedical imaging.
• To enhance the effectiveness of active learning by better leveraging both labeled and unlabeled data.
• To create an end-to-end active learning framework for connectomics tasks.

Main Methods:

• A two-stream active learning approach is proposed, incorporating both supervised and unsupervised feature extractors.
• The unsupervised stream is optimized on all raw images to capture diverse features, improving generalization.
• An end-to-end framework was developed for 3D synapse detection and mitochondria segmentation in connectomics.

Main Results:

• The proposed method significantly outperforms state-of-the-art approaches in 3D synapse detection (3.1% improvement) and mitochondria segmentation (3.8% improvement) on a newly curated connectomics dataset.
• The framework demonstrated superior performance on the CIFAR-10 image classification dataset under limited annotation budgets.
• A large-scale connectomics dataset with dense synapse and mitochondria annotation was curated.

Conclusions:

• The two-stream active learning method effectively addresses the challenge of limited labeled data in large-scale biomedical vision tasks.
• This approach enhances the performance of deep learning models by better utilizing unlabeled data for feature extraction.
• The developed framework and curated dataset advance the field of connectomics research and biomedical image analysis.