Nariman Majdi Nasab PhD, Presenter: Nothing to Disclose

Holman Page McAdams MD, Abstract Co-Author: Nothing to Disclose

Devon James Godfrey PhD, Abstract Co-Author: Nothing to Disclose

James Talmage Dobbins PhD, Abstract Co-Author: Nothing to Disclose

Digital tomosynthesis shows promise for improved lung nodule detection compared to chest radiography. It may also improve performance of CAD algorithms for lung nodule detection by eliminating structure overlap that causes many false positives (FP). We report our initial experience applying a CAD algorithm to tomosynthesis images acquired as part of an NIH-sponsored clinical trial.

An algorithm (CAD-T) was developed to identify nodules in tomosynthesis images and was applied to 5 subjects enrolled in the ongoing trial. The algorithm consists of (1) image filtration using difference of Gaussian, Wiener, and linearization filters, (2) image segmentation using Markov random field modeling, and (3) feature extraction using 14 morphological and textural features. Enrolled subjects were recruited from among patients undergoing CT to follow-up lung nodules. They received digital tomosynthesis using a commercial CsI/a-Si flat-panel detector and a custom-built tube mover. Acquisition parameters in the 5 subjects were: 71 projection images, 20° tube angle, 10-second breath-hold. Reconstruction parameters were: MITS algorithm, 69 planes, 5mm plane spacing, sliding 7-plane averaging to reduce low-contrast artifacts. Patient exposure was equivalent to a lateral film-screen radiograph. An experienced chest radiologist marked all true nodules on tomosynthesis images, using CT data for localization. Nodules found by CAD-T were marked as true positive (TP) if within 2 mm of the nodule center identified by the chest radiologist.

There were 12 nodules in the 5 subjects, ranging in size from 4 to 21mm (mean=10.4mm). The CAD-T method detected 11 and missed 1 nodule, for an average sensitivity of 92%. There were an average of 7 FP marks per subject. FP marks were mainly due to end-on vessels seen in cross section.

Our CAD algorithm applied to tomosynthesis images showed excellent sensitivity and comparable specificity for lung nodule detection to other CAD methods applied to projection chest radiographs. These preliminary results are being augmented by additional cases as they are accrued. Methods to reduce false positives due to end-on vessels are being developed.

N.M.,H.P.M.,D.J.G.,J.T.D.: This research was supported in part by a grant from NIH (R01 CA80490). None of the authors has a direct financial interest. The flat-panel detector used in the study was obtained through a research agreement with GE Medical Systems.

Majdi Nasab, N, McAdams, H, Godfrey, D, Dobbins, J, Automated Detection of Lung Nodules on Digital Tomosynthesis Images Using a Textural and Morphological CAD Approach: Early Results.  Radiological Society of North America 2005 Scientific Assembly and Annual Meeting, November 27 - December 2, 2005 ,Chicago IL. http://archive.rsna.org/2005/4414640.html