Uncertainty-based Active Learning by Bayesian U-Net for Multi-label Cone-beam CT Segmentation.

Area of Science:

• Biomedical image analysis
• Artificial intelligence in dentistry
• Medical imaging and diagnostics

Background:

• Training AI for biomedical image analysis requires large annotated datasets.
• This study focuses on improving AI model training for periapical lesion detection in CBCTs.
• Limited datasets pose a challenge for developing accurate AI models in dentistry.

Purpose of the Study:

• To assess the efficacy of Active Learning (AL) strategies for training AI models.
• To improve multilabel segmentation and detection of periapical lesions in CBCTs.
• To evaluate AL strategies using limited annotated data.

Main Methods:

• Utilized limited field-of-view CBCT volumes (n=20) for training.
• Employed Bayesian U-Net with two AL functions: Bayesian Active Learning by Disagreement (BALD) and Max_Entropy (ME).
• Compared AL strategies against a non-AL Bayesian U-Net benchmark, evaluating segmentation accuracy (Dice indices) and lesion detection.

Main Results:

• After 8 AL iterations, lesion detection sensitivity reached 84.0% for BALD and 76.0% for ME, significantly outperforming the non-AL Bayesian U-Net (32.0%).
• Mean Dice indices for all labels were higher with AL strategies (0.703 ± 0.166 for ME) compared to the non-AL benchmark (0.680 ± 0.155).
• The Dice index for 'Lesion' was 0.504 for BALD and 0.501 for ME after 8 AL iterations, significantly higher than the non-AL model's maximum of 0.288.

Conclusions:

• Active Learning strategies, particularly BALD and ME, enhance AI model accuracy for segmentation and lesion detection in CBCTs.
• AL, leveraging uncertainty quantification, effectively reduces the need for extensive data annotation in dental AI.
• This approach holds promise for more efficient development of AI tools in biomedical image analysis for dentistry.