Fully Automatic Volumetric Segmentation of Pulmonary Nodules: Evaluation using the Complete LIDC/IDRI Database

B. Lassen, J. Kuhnigk, C. Jacobs, E. van Rikxoort and B. van Ginneken

Annual Meeting of the Radiological Society of North America 2014.

PURPOSE In the publicly available LIDC/IDRI database, all nodules larger than 3mm have been manually segmented by four expert thoracic radiologists. This provides a unique opportunity for large scale validation. We report the performance of our automatic method to segment pulmonary nodules and compare this to inter-reader variability. METHOD AND MATERIALS We developed an automatic nodule segmentation method which is initialized by region growing from a seed point in the nodule. Thresholds for region growing are automatically determined from histogram analysis. A circumscribing ellipsoid is approximated to separate nodules from the chest wall. Finally, through a combination of connected component analysis and morphological operations vasculature attached to the nodule is removed. To evaluate our automatic method, it was applied four times using a random seed point in each nodule in the LIDC/IDRI database that contains 1,018 chest CT scans from 1,000 patients, acquired at seven different institutions with a wide variety of scanners and imaging protocols. In these scans, 928 nodules were manually segmented independently by four radiologists by drawing contours on each axial section containing the nodule. The DICE overlap between the resulting automatic outline and the three other manual segmentations was computed. Similarly, each manual segmentation was compared to the three other manual outlines. We report statistics of the averaged DICE coefficients. RESULTS We achieved excellent agreement between our automatic and manual segmentation results. Mean DICE was 0.75 A,A+- 0.16 for the automatic method and 0.77 A,A+- 0.09 for the inter-observer agreement. The first quartile, median, and third quartile for the automatic method were 0.71, 0.79. 0.84, respectively. For the manual outlines, these statistics were 0.73, 0.79, 0.83. CONCLUSION Automated nodule segmentation is feasible in CT scans obtained with varying acquisition parameters with a performance close to manual outlining by expert thoracic radiologists. CLINICAL RELEVANCE/APPLICATION Automatic volumetric nodule segmentation is a robust, efficient and highly effective technique for the analysis of pulmonary nodules in CT data.