Computer-Aided Detection of Ground Glass Nodules in Thoracic CT Images

C. Jacobs, E. Scholten, S. Saur, T. Twellmann, P. de Jong and B. van Ginneken

Annual Meeting of the Radiological Society of North America 2011.

PURPOSE : Ground glass nodules (GGNs) occur less frequent in computed tomography (CT) scans than solid nodules but have a much higher chance of being malignant. Therefore, accurate detection of these nodules is highly important. A complete system for computer-aided detection (CAD) of GGNs is presented and evaluated on an independent data set to determine its potential to be implemented in a clinical setting. METHOD AND MATERIALS: Data for this study was collected from 10,000 low dose chest CT scans (16x0.75mm, 120-140 kVp, 30 mAs) from a large lung cancer screening trial. All examinations in which at least one GGN was annotated were collected. This resulted in 140 scans. This set is extended with 60 randomly selected scans in which two radiologists confirmed the absence of GGNs. Training and optimization of the CAD system was performed using 67 scans each containing one or more GGNs, 91 in total. The CAD system was independently tested using FROC analysis on the remaining 133 scans, 73 with and 60 without GGNs, containing 85 GGNs in total. The detection pipeline is initiated with a robust lung and airway segmentation. The CAD system extracts ground glass voxels by applying a double-threshold density mask (-750 to -300 HU) to the lung regions. A morphological opening operation, connected component analysis and a size criterion define GGN candidates. Classification of candidates is accomplished using a two-stage classification process. First, a linear discriminant classifier using 2 shape and 2 intensity features is employed. Next, a GentleBoost classifier is applied to the remaining candidates using a rich set of 161 features that describe the intensity, shape and spatial position of candidates relative to the airways and lung boundary. RESULTS: Candidate extraction resulted in 560 ? 306 candidate regions per scan. On average, 32% of the candidate regions remained after the first classification stage. The CAD system achieved a sensitivity of 60% and 74% at 0.25 and 1 false positive detections per scan, respectively. Most false negatives were missed in the candidate detection stage.CONCLUSION: Computer-aided detection of ground glass nodules is feasible with high sensitivity at a low false positive rate. CLINICAL RELEVANCE/APPLICATION: Ground glass nodules are easily missed in chest CT scans. Computer-aided detection can reach high sensitivity at low false positive rates and is therefore a potentially useful aid for radiologists.