Jorge Juan Suarez-Cuenca, Presenter: Nothing to Disclose

Qiang Li PhD, Abstract Co-Author: Patent agreement, General Electric Company Patent agreement, Hologic, Inc Patent agreement, Riverain Medical Patent agreement, MEDIAN Technologies Patent agreement, Mitsubishi Corporation

Miguel Souto MSC, Abstract Co-Author: Nothing to Disclose

Pablo G. Tahoces PhD, Abstract Co-Author: Nothing to Disclose

Martine J. Remy-Jardin MD, PhD, Abstract Co-Author: Research grant, Siemens AG

Jacques Remy MD, Abstract Co-Author: Research Consultant, Siemens AG

CAD systems for lung nodule detection have been developed over the past few years, but these systems were usually trained and tested with the same databases by employing cross validation or leave one out methods. A “good” CAD evaluated with a database by such methods does not necessary achieve good performance for a different database. The purpose of this study was to evaluate the generalizability of our CAD system by using two independent databases.

Two databases of chest CT images were collected from two institutions in Spain and France. Database 1 included 20 CT scans with 49 nodules, and was reconstructed at 1.25 mm slice thickness with a kernel B90s from a 6-slice CT scanner SOMATOM EMOTION. Acquisition parameters include 6x1.0 mm collimation, 130 KVp, and 70 mAs. Database 2 included 43 cases with 83 pulmonary nodules and was reconstructed at 1.0 mm slice thickness with a B50f kernel from a 64-slice dual source CT scanner SOMATOM DEFINITION. Acquisition parameters included 64x0.6 mm collimation, 100-120 KVp and 100-110 mAs. The 132 nodules included in the study ranged from 4 to 30 mm in diameter, and were localized by three experienced chest radiologists. Our CAD system automatically identified suspicious candidates, extracted gray-level morphological features for candidates, used iris filter to characterize candidates, and employed a quadratic classifier to reduce the number of false positives. The system was trained with Database 1 and tested with Database 2, and vice verse. FROC curves, sensitivity and number of false-positive per scan, were employed to evaluate the performance of the CAD system.

When the CAD was trained with Database 2 and tested with Database 1, it achieved a sensitivity of 79.6% and a FP rate of 5 per CT scan. When the CAD was trained with Database 1 and tested with Database 2, it yielded a sensitivity of 79.5% at a FP rate of 5 per CT scan.

Our CAD system achieved a sensitivity of 79.5%, with 5 false positives per examination when it was tested with independent databases. The system is well generalizable since it can be applied to databases obtained from different machines with similar performance.

A generalizable computer-aided diagnostic (CAD) system maintains high performance for CT scans acquired from different hospitals, and can thus find wide applications in clinical practice.

Suarez-Cuenca, J, Li, Q, Souto, M, Tahoces, P, Remy-Jardin, M, Remy, J, Generalizability of CAD System for Lung Nodule Detection in CT: Evaluation of CAD with Independent Databases.  Radiological Society of North America 2010 Scientific Assembly and Annual Meeting, November 28 - December 3, 2010 ,Chicago IL. http://archive.rsna.org/2010/9007053.html