Evaluation of a Deep Active Learning Model for the Segmentation of Canine Thoracic Radiographs

Area of Science:

• Veterinary radiology
• Medical imaging analysis
• Artificial intelligence in veterinary medicine

Background:

• Image segmentation is crucial for veterinary radiology, but manual methods are time-consuming and vary in accuracy.
• There's a need to evaluate AI segmentation tools for veterinary medical imaging to assess their impact on accuracy and efficiency.
• It remains unclear if novice users can achieve expert-level accuracy in radiograph segmentation.

Purpose of the Study:

• To evaluate an AI segmentation tool's performance in enhancing accuracy and reducing segmentation time for canine radiographs.
• To compare AI-assisted semiautomated segmentation with manual segmentation across novice, intermediate, and expert users.
• To assess the impact of AI on inter-user variability and segmentation consistency.

Main Methods:

• An AI model was trained on 50 canine thoracic radiographs.
• A custom software allowed for both AI-assisted semiautomated and manual segmentation.
• Intersection over Union (IoU) scores and Hausdorff distance were used to evaluate segmentation accuracy.
• Intraobserver intraclass correlation coefficients (ICC) assessed segmentation consistency.

Main Results:

• Semiautomated AI segmentation yielded higher IoU scores for abdomen, heart, and spinous process compared to manual segmentation.
• Hausdorff distance was significantly lower with the semiautomated method, indicating greater accuracy.
• Intraobserver ICC was substantially higher for the semiautomated method (0.81) than the manual method (0.36).

Conclusions:

• The AI-assisted semiautomated tool significantly improved the accuracy and consistency of canine thoracic radiograph segmentation.
• The AI tool reduced segmentation time and user variation, benefiting users of all expertise levels.
• AI segmentation shows promise for advancing veterinary radiology by improving efficiency and diagnostic support.