UCLA researchers have developed an automated image processing system that segments lung regions in CT scans, identifies candidate high-intensity structures (like nodules), and classifies lung nodules (including solid and ground glass) using geometric analysis. The system aims to reduce false positives while maintaining high sensitivity in lung cancer screening workflows.
Lung cancer screening via computed tomography (CT) is effective but time-consuming and subject to radiologist variability. Detecting small lung nodules—solid or ground-glass—amid surrounding anatomical structures (airways, vessels) is challenging. False positives (blood vessels, imaging artifacts) burden clinical workflow and may lead to unnecessary follow-ups or diagnostic stress. There is a need for reliable, automated tools that assist radiologists by segmenting lung regions, filtering out non-nodule structures, and flagging real nodules with high specificity and maintained sensitivity.
The system first segments lung region(s) from the chest CT volume using intensity thresholding and morphological / region-growing methods to exclude the chest wall and non-lung tissues.
It then applies intensity thresholding to identify high-intensity voxels (potential nodules or anatomical structures) within lung regions.
A Euclidean distance map is computed on the voxel mask of high-intensity regions, followed by watershed segmentation to break the mask into sub-regions.
For each sub-region, a seed point (voxel of maximal distance) is selected; candidate regions are then grown from those seed points. Geometric features (size, sphericity, volume, etc.) are extracted to distinguish nodules (solid or ground glass) from vessel/tubular or non-nodule structures.
The method includes thresholds (e.g. diameter ≥ 4 mm or twice slice thickness) to define significant nodules and reduce false positives.
Automated processing reduces manual workload for radiologists.
Effective detection of both solid and ground glass nodules.
Lower false positive rate by using shape and geometric discrimination (sphericity, distinguishing vessel-like vs nodule-like morphology).
Ability to flag nodules with clinically relevant size thresholds, improving consistency.
Applicable to full CT volumes—scalable for screening programs.
Lung cancer screening workflows to assist radiologists in early detection.
Triage tools in diagnostic radiology to flag suspicious nodules for further review.
Longitudinal monitoring of patients with lung nodules (follow-up, growth detection).
Integration into CAD (computer-aided detection/diagnosis) software for hospitals or imaging centers.
Research studies analyzing lung nodule prevalence, growth, morphology in patient cohorts.
US 9,418,420 B2 — System and method for automated detection of lung nodules in medical images Google Patents
PCT publication WO2014/042678 A1 in the same patent family. Google Patents+1