Interpolation-split: a data-centric deep learning approach with big interpolated data to boost airway segmentation...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Airway segmentation is crucial for diagnosing and characterizing chronic respiratory diseases.
• Accurate segmentation of the entire airway tree is challenging due to variations in intensity, scale, and shape.
• Existing methods often result in under- or oversegmentation, requiring manual correction, while deep learning methods demand high computational resources.

Purpose of the Study:

• To develop a data-centric deep learning technique for accurate and efficient airway tree segmentation.
• To address the limitations of existing methods, including manual intervention and high computational costs.
• To improve the estimation of disease extent and severity in chronic respiratory conditions through enhanced segmentation.

Main Methods:

• Proposed a novel data-centric deep learning technique named Interpolation-Split.
• Utilized interpolation and image splitting to enhance data quality and usefulness.
• Implemented an ensemble learning strategy to aggregate segmented airway components across different scales.

Main Results:

• Achieved average segmentation performance (Dice Similarity Coefficient) of 90.55%, 89.52%, and 85.80% on different baseline models.
• Outperformed baseline models by an average of 2.89% to 3.87%.
• Demonstrated significant performance gains, up to 14.11%, with low RAM and GPU memory usage, proving GPU memory efficiency and flexibility.

Conclusions:

• The Interpolation-Split technique significantly improves airway tree segmentation performance.
• The method is computationally efficient, requiring low memory, and can be deployed on various 2D deep learning models.
• This approach offers a promising solution for automated and accurate airway segmentation in clinical settings.