An approach for reducing the error rate in automated lung segmentation.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Radiology

Background:

• Accurate lung segmentation is critical for quantitative analysis in large-scale lung CT studies.
• Existing segmentation methods, like region growing and model-based approaches, face challenges with robustness and failure rates.

Purpose of the Study:

• To develop and evaluate a fusion method combining segmentations from two different approaches.
• To achieve lung segmentations with a lower failure rate than individual methods.

Main Methods:

• Utilized lung segmentations from region growing and model-based methods as input.
• Developed a trained classification system to selectively combine input segmentations based on comparison.
• Evaluated the fusion method on 204 diverse CT scans of normal and diseased lungs.

Main Results:

• The fusion approach achieved a high Dice coefficient (0.9855±0.0106), significantly outperforming individual methods.
• Demonstrated a substantially lower failure rate; e.g., 6.13% failure for Dice > 0.97, versus 18.14% and 15.69% for the input methods.
• Reported results on the LOLA11 challenge test set for inter-method comparison.

Conclusions:

• The proposed fusion method enhances lung segmentation quality and robustness.
• Improved segmentation accuracy is vital for reliable subsequent quantitative analysis of lung imaging data.
• This technique offers a more reliable solution for processing large datasets in multi-center studies.