Automatic Detection of Large Pulmonary Nodules in Thoracic CT Images

A. Setio, J. Gelderblom, C. Jacobs and B. van Ginneken

Annual Meeting of the Radiological Society of North America 2014.

PURPOSE: Existing computer-aided detection (CAD) systems excel at finding small nodules but often fail to detect the much rarer larger nodules. However, these large nodules are highly suspicious for being cancer. Therefore, we developed a CAD system specifically designed to detect large nodules. METHOD AND MATERIALS: Data from the publicly available LIDC/IDRI database was used. CT scans with section thickness over 2.5 mm were excluded. We selected all scans in which at least one of the four radiologists who read each case, annotated a solid nodule larger than 10 mm. The detection pipeline is initiated by a three-dimensional lung segmentation algorithm. Large nodules attached to the pleural wall are often excluded in this segmentation. Therefore, a rolling-ball algorithm was applied to the lung segmentation to include these nodules. The detection of nodule candidates was performed using a cascade of double threshold on intensity, followed by morphological operation. Connected component analysis was subsequently applied to get initial nodule candidates. The segmentation of the initial candidates was refined using a previously published nodule segmentation method. For each candidate, a total of nine intensity and shape features were extracted. A Support Vector Machine (SVM) classifier with a radial basis function was used to classify nodule candidates and performance was evaluated using a 10-fold cross-validation scheme. CAD marks on nodules annotated by all four radiologists were counted as true positives. CAD marks on nodules annotated by less than four radiologists were ignored in the analysis. Other CAD marks were considered false positives. RESULTS: In 271 scans, 208 large nodules were annotated by all four radiologists. The candidate detection stage detected 98.6% (205/208) of the large nodules, with an average of 44.6 false positives per scan. After classification, the CAD system achieved a sensitivity of 95.7% (199/208) and 84.6% (176/208) at 7.5 and 1.0 false positives per scan, respectively. CONCLUSION: A dedicated CAD system for large pulmonary nodules can identify the vast majority of highly suspicious lesions in thoracic CT scans with a small number of false positives. CLINICAL RELEVANCE/APPLICATION: As computers start to gain a more important role in CT lung cancer screening, it is vital that CAD reaches a high sensitivity in the detection of large nodules, which are likely to be cancer.