Meeting Abstracts
» Double Reading of Chest CT Scans: Comparison of Reader-Reader Pairing versus Computer-aided Detection (CAD)–Reader Pairing for the Detection of Lung Nodules

PURPOSE: To determine the potential improvement in nodule detectability by radiologists when assisted by CAD compared to double reading by radiologist pairs.

METHODS AND MATERIALS: Three faculty radiologists individually reviewed MDCT scans (4 x 1.25 mm, 6.0 pitch) from 21 patients suspected of having pulmonary nodules. The radiologists identified nodule candidates and scored them on a 5-point confidence scale (5 = definite, 3 = possible, 1 = very unlikely). CAD was performed independently on the 21 scans. A consensus panel of two thoracic radiologists established the ground truth of 100 ≥ 3 mm nodules in 14 and no nodules in 7 of the 21 patients. Nodule candidates for radiologist-radiologist or radiologist-CAD pairs were combined using the “OR” rule, where a positive result by at least one reader was considered positive for the double reading.

RESULTS: The three radiologists achieved mean sensitivities for detection of ≥3mm pulmonary nodules of 73, 80, and 85% (mean ± sd = 79  3.0%) and a mean of 2.2, 2.2, and 3.9 false positive (FP) results per scan, respectively for the highest three confidence levels. Mean radiologist interpretation times were 7.4, 4.7, and 9.8 minutes, respectively. CAD achieved peak sensitivities of 96 and 89% with an average of 67 and 29 FP per scan, respectively. When considering the three highest confidence scores, radiologist–radiologist pairings resulted in a sensitivity increase of 4-16% (mean ± sd =10.3  3.0%) to 88-92%. CAD–radiologist pairings improved radiologist sensitivity by 8-17% (mean ± sd = 19.0  3.1 %) to 90-93% at a CAD sensitivity threshold that resulted in 10.0 FP/scan. When the CAD sensitivity threshold was reduced to allow an average of 4.9 FP/scan (similar to radiologist performance), CAD-radiologist pairings resulted in an 5-15% (mean ± sd = 11.0 ± 4.7%) to 87-90%.

CONCLUSION: Substantial variations in radiologist performance existed when identifying pulmonary nodules. The addition of CAD to the radiologist reading narrowed the range of sensitivities from 12 to 3%, due to greater improvement of less sensitive radiologists by CAD. CAD may improve radiologist sensitivity to a level comparable to double readings by radiologist pairs and possibly higher. The degree of improvement is directly related to the number of FP results generated by CAD that the radiologist must exclude. However, radiologist screening of even a modest number of FP results (5–10 per scan) may improve radiologist sensitivity substantially.

This page last modified on 3 February 2004 by Society Webmaster.