Automated detection of lung nodules in multidetector CT: influence of different reconstruction protocols on performance of a software prototype.

PURPOSE

To evaluate the accuracy of software for computer-aided detection (CAD) of lung nodules using different reconstruction slice thickness protocols in multidetector CT.

MATERIALS AND METHODS

Raw image data sets for 15 patients who had undergone 16-row multidetector CT (MDCT) for known pulmonary nodules were reconstructed at a reconstruction thickness of 5.0, 2.0 and 1.0 mm with a reconstruction increment of 1.5, 1.0 and 0.5 mm, respectively. The "Nodule Enhanced Viewing" (NEV) tool of LungCare for computer-aided detection of lung nodules was applied to the reconstructed images. The reconstructed images were also blinded and then evaluated by 2 radiologists (A and B). Data from the evaluating radiologists and CAD was then compared to an independent reference standard established using the consensus of 2 independent experienced chest radiologists. The eligible nodules were grouped according to their size (diameter > 10, 5 - 10, < 5 mm) for assessment. Statistical analysis was performed using the receiver operating characteristic (ROC) curve analysis, t-test and two-rater Cohen's Kappa co-efficient.

RESULTS

A total of 103 nodules were included in the reference standard by the consensus panel. The performance of CAD was marginally lower than that of readers at a 5.0-mm reconstruction thickness (AUC = 0.522, 0.517 and 0.497 for A, B and CAD, respectively). In the case of 2.0-mm reconstruction slices, the performance of CAD was better than that of the readers (AUC = 0.524, 0.524 and 0.614 for A, B and CAD, respectively). CAD was found to be significantly superior to radiologists in the case of 1.0-mm reconstruction slices (AUC = 0.537, 0.531 and 0.675 for A, B and CAD, respectively). The sensitivity at a reconstruction thickness of 1.0 mm was determined to be 66.99 %, 68.93 % and 80.58 % for A, B and CAD, respectively. The time required for detection was shortest for CAD at reconstruction slices of 1.0 mm (mean t = 4 min). The performance of radiologists was greatly enhanced when using CAD: sensitivity 91.26 % and 94.17 % for CAD+A and CAD+B, respectively (AUC = 0.889 and 0.917). CAD was most advantageous in the detection of nodules < 10 mm.

CONCLUSION

At a 1.0-mm reconstruction thickness, CAD's ability to detect nodules < 10 mm is superior to that of radiologists and its relatively short evaluation time makes it a viable second reader.