Ori Hay MS, PRESENTER: Nothing to Disclose

Abstract: HTML Purpose: Manual, volume segmentation of lung nodules seen on CT images is a tedious and time-consuming process. In this study, an automated tool for lung nodule segmentation is evaluated on phantom data for standard and low dose scans. Methods and Materials: Lung phantoms containing 34 nodule-like objects (14 spheres and 20 cylinders) were scanned on a Philips MX8000 IDT CT scanner using various scanning and reconstruction parameters including mAs (20-300 mAs), slice width, slice increment, filtering, etc. Nodule objects were spheres and cylinders of various sizes (17 square mm to 1700 square mm) in air or attached to phantom lung wall. The nodule objects were segmented using automated segmentation tool (Lung nodule assessment and comparison package, MxView, Philips Medical Systems) by placing a seed on the nodule. The segmentation is independent of image window or level, repeatable, and estimates the volume using 3D data. In this study, comparisons were made of known nodule volume versus volume from automated segmentation and segmentation performance at various mAs. The evaluation consisted of automated segmentation of 34 different objects of various sizes (spheres with diameters 3,4,6,8,10,12 mm and cylinders with diameters and height pairs (in mm) of (3,5), (4,5), (4,7.5), (6,7.5), (8,10), (10,15), (12,15)). Scans were acquired at 5 acquisition protocols and at 25,35,50,150,200, and 250 mAs. Results: The segmentation error mean was 5.6% with standard deviation (STD) of 3.6% for cylinders and 7.1% with STD of 5.7% for spheres for standard dose scans (200-250 mAs). There is a strong correlation between automated segmentation and ground truth with R2 > 0.99 for both spheres and cylinders. For low dose scans (20-30 mAs) segmentation error mean was 5.4% with STD of 4.0% for cylinders and 7.5% with STD of 4.0% for spheres. There is a strong correlation between automated segmentation and ground truth with R2 > 0.95 for both spheres and cylinders. The mean difference in volume between standard and low dose scan was 5.1% with STD of 4.6% for cylinders and 8.5% with STD of 7.2% for spheres. There is a strong correlation between standard and low dose automated segmentation with R2 > 0.98 for both spheres and cylinders. Conclusion: Our results suggest that automatic nodule segmentation is accurate for volume calculation and robust even for low-dose, small nodules, and nodules attached to lung wall. (O.H., D.S., Y.S. are employees of Philips Medical Systems. R.W. is an employee of Philips Research.) Questions about this event email: ori.hay@philips.com

Hay MS, O, Evaluation of Automatic Volumetric Segmentation of Lung Nodules in Standard and Low Dose CT Scans.  Radiological Society of North America 2003 Scientific Assembly and Annual Meeting, November 30 - December 5, 2003 ,Chicago IL. http://archive.rsna.org/2003/3107117.html