Abstract
Accurate segmentation of lesions is an essential process in CAD systems for differential diagnosis. However, segmentation of ground glass opacity (GGO) nodules, most of which are likely to be malignant, is a difficult task due to their poorly defined margins. We have developed a new segmentation algorithm for GGO nodules that enhances lesions by applying a boosting algorithm based on CT value features and extracts GGO nodules by graph cuts. Experiments to validate the proposed algorithm were conducted using data for 100 GGO nodules acquired at Kyoto University Hospital, and the Jaccard index between the extracted region and the true region was computed by a 10-fold cross-validation test. The Jaccard index of conventional graph cuts using CT data only was 40.7%, and the binarization results for lesions enhanced by a boosting algorithm showed a Jaccard index of 67.3%. In contrast, the proposed method, which combines enhancement processing with graph cuts, achieved a Jaccard index of 72.2%, which is significantly higher than the above performance indices.