Active learning for data efficient semantic segmentation of canine bones in radiographs.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthopedics

Background:

• X-ray bone semantic segmentation is vital in medical imaging for precise analysis.
• Deep learning models offer high accuracy but demand extensive annotated datasets.
• Pixel-wise labeling for segmentation is time-consuming, especially for complex anatomy like hip joints.

Purpose of the Study:

• To investigate active learning strategies for efficient X-ray bone semantic segmentation.
• To compare the effectiveness of uncertainty and diversity-based queries in active learning.
• To reduce the amount of annotated data required for high-performance segmentation models.

Main Methods:

• Implementation of active learning framework for semantic segmentation.
• Comparison of different query strategies: uncertainty sampling and diversity sampling.
• Evaluation of model performance based on varying percentages of annotated data.

Main Results:

• Active learning significantly reduces the need for annotated data.
• Proposed query methods achieve state-of-the-art performance.
• The approach utilizes only 81.02% of the data required by traditional methods.

Conclusions:

• Active learning is an effective strategy to minimize data requirements in X-ray bone segmentation.
• Uncertainty and diversity-based queries improve model efficiency.
• This method offers a practical solution for data scarcity in medical image analysis.