A fully automatic segmentation algorithm for CT lung images based on random forest.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Radiology

Background:

• Accurate lung segmentation in computed tomography (CT) images is challenging due to factors like juxta-pleural nodules, pulmonary vessels, and image noise.
• Existing methods often struggle to achieve high precision in complex thoracic CT scans.

Purpose of the Study:

• To develop a novel hybrid automated algorithm based on random forest for precise lung segmentation in CT images.
• To effectively eliminate the influence of negative factors and enhance the accuracy of lung segmentation.

Main Methods:

• A five-step process involving image preprocessing (denoising, decomposition), superpixel segmentation, and random forest classification.
• Refinement steps include trachea elimination, lung separation using contextual information, and contour correction with corner detection.

Main Results:

• The algorithm achieved high accuracy on CT images of patients with interstitial lung diseases, demonstrating a Jaccard's index of 0.9638 and a Dice similarity coefficient of 0.9867.
• Outperformed conventional methods by an average of 7.7% in Dice similarity coefficient and surpassed Deep Learning methods by 1%.

Conclusions:

• The proposed hybrid algorithm provides a fully automatic and high-performance solution for lung segmentation from CT images.
• This method significantly assists in lung disease detection within computer-aided detection systems.