"Computer aided detection of lung nodules in chest radiographs: novel approaches to improve reader performance"

S. Schalekamp, B. van Ginneken, C. Schaefer-Prokop and N. Karssemeijer

in: MIPS, 2013

Abstract

RATIONALE: To investigate new methods of using a computer aided detection (CAD) system for the detection of lung nodules in chest radiographs. CAD systems for the detection of lung nodules may be clinically used but their standalone performance is still low in comparison with radiologists. Since CAD is limited by relatively high false positive rates, the way these systems are currently used may not be optimal. In our study we investigate two alternatives: Use of CAD as interactive decision support and CAD as an independent second reader. METHODS: We selected 300 frontal and lateral digital chest radiographs, including 111 with a solitary pulmonary nodule. All chest radiographs were analyzed by a commercially available CAD system (ClearRead +Detect 5.2, Riverain Technologies, Miamisburg, Ohio) which provided lesion contours of suspicious regions, accompanied by a probability score. The CAD system was used in an interactive manner. CAD marks and their suspiciousness score remained hidden unless their location was queried by the radiologist. Twelve radiologists read the CXRs without CAD and with the interactive CAD in two reading sessions. AFROC MRMC analysis was used to measure detection performance. Partial areas under the curve in a FPF range between 0 and 0.2 were used to compare reader performance. Besides, results of a weighted independent combination of CAD scores and reader scores, at the location of reader findings, were evaluated. RESULTS: Average partial area under the curve for radiologists without CAD was 0.127. No improvement was seen for radiologists with use of the interactive CAD (pAUC=0.127, p=0.88). Independent combination of reader scores with CAD significantly improved performance (pAUC=0.135,p=0.007). CONCLUSION: Though interactive use of CAD did not improve reader performance for the detection of lung nodules in chest radiographs, CAD has potential as decision support, since a simple weighted combination of reader scores with CAD scores significantly improved performance. Lack of confidence in CAD in the interactive sessions may explain these results.