Cavity contour segmentation in chest radiographs using supervised learning and dynamic programming.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Pulmonary disease diagnostics

Background:

• Chest radiography is crucial for monitoring tuberculosis (TB) treatment efficacy.
• Cavity size in pulmonary TB predicts disease severity and relapse risk.
• Accurate segmentation of cavities is essential for reliable monitoring.

Purpose of the Study:

• To develop and evaluate an automated method for segmenting cavities in chest radiographs.
• To assess the performance of the automated segmentation against expert manual delineations.

Main Methods:

• A two-stage approach using supervised learning for pixel classification and dynamic programming for contour tracing.
• Utilized texture and radial features for training pixel classifiers (kNN, LDA, GB, RF).
• Evaluated on 100 chest radiographs with 126 cavities, comparing against radiologist-drawn segmentations.

Main Results:

• Achieved a median Jaccard overlap of 0.81 with reference segmentations, comparable to inter-radiologist variability (0.85).
• Mean contour and Hausdorff distances were 2.48 ± 2.19 mm and 8.32 ± 5.66 mm, respectively.
• 84% of automatic segmentations were rated as 'excellent' or 'adequate' by expert readers.

Conclusions:

• The proposed automated cavity segmentation method demonstrates high accuracy and good overlap with expert segmentations.
• The technique's performance approaches that of experienced radiologists.
• This automated approach can support tuberculosis treatment monitoring, particularly in resource-limited settings.