Robust semi-automatic segmentation of pulmonary subsolid nodules in chest computed tomography scans

Area of Science:

• Medical Imaging
• Radiology
• Computational Pathology

Background:

• Malignancy detection in lung nodules typically relies on diameter changes in follow-up scans.
• Nodule volume or mass changes are more significant indicators of malignancy than diameter.
• Subsolid lung nodules have a higher malignancy rate than solid nodules, yet segmentation methods are less developed.

Purpose of the Study:

• To present a fast, semi-automatic method for segmenting subsolid lung nodules.
• To reduce the time-consuming nature of manual nodule volume segmentation.
• To decrease inter-observer variability in subsolid nodule segmentation for earlier and more accurate lung cancer detection.

Main Methods:

• A semi-automatic segmentation approach requiring minimal user input (a single user-drawn stroke).
• Combines threshold-based region growing, intensity analysis, connected component analysis, convex hull calculation, and morphological operations.
• Specifically designed to handle chest wall removal and detachment of attached vessels.

Main Results:

• Achieved a Jaccard index of 0.52/0.50 on the LIDC/IDRI database (Dataset 1/Dataset 2), comparable to manual inter-observer agreement (0.54/0.58).
• Demonstrated improved inter-observer agreement with the proposed method (Jaccard index of 0.74/0.74) when using different input strokes.
• The method provides satisfactory segmentation with minimal observer effort and time.

Conclusions:

• The developed method offers a fast and efficient solution for subsolid lung nodule segmentation.
• It significantly reduces inter-observer variability, potentially improving clinical routine diagnosis.
• This technique aids in earlier detection of lung cancer by facilitating more consistent and rapid analysis of subsolid nodules.