Colin Craig McCulloch PhD, Abstract Co-Author: Nothing to Disclose

Claudia Ingrid Henschke MD, Abstract Co-Author: Nothing to Disclose

Ella Annabelle Kazerooni MD, Abstract Co-Author: Nothing to Disclose

Joseph Ken Leader, Abstract Co-Author: Nothing to Disclose

Wesley Turner PhD, Presenter: Nothing to Disclose

Ricardo Avila, Abstract Co-Author: Nothing to Disclose

To measure the potential for a CAD algorithm to improve radiologist sensitivity for solid and sub-solid non-calcified nodules (NCN) in low dose CT exams.

A 9 reader 3 institution study is underway to assess the variability of radiological readings of low dose thin slice CT lung cancer screening exams (20-60 mAs, 2.5mm, acquired on LightSpeed scanners (GE Healthcare). To date, a total of 492 exams have been read 3 times, where the 3 readers at each institution read an institution-specific exam set: 292 exams at the first site, 111 at the second, and 89 at the third. All readers, board-certified radiologists, read the exams independently using the GE Advanced Lung Analysis package. A prototype model-based CAD algorithm developed at GE Global Research was used to evaluate potential radiological sensitivity improvements stratified by NCN type. Sensitivity was defined in this paper with respect to majority-detected NCNs that were considered suspicious for lung cancer and larger than 4mm. Sensitivity improvements were inferred by combining radiological readings with CAD results, assuming readers would not reject any CAD true positives.

Of the 642 majority-detected NCNs, 497 (77%) were solid, 79 (12%) were part-solid, and 66 (10%) were nonsolid. Of the 280 (44%) NCNs not identified by one reader, 218 (78%) were solid, 34 (12%) were part-solid, and 28 (10%) were nonsolid. Average radiological sensitivity was 87%. Sensitivity for solid, part-solid, and nonsolid NCNs was 87%, 89%, and 86%, respectively.The CAD algorithm detected 119 of the 280 NCNs not identified by one reader. Using this algorithm, average radiological sensitivity for the NCNs would have increased to 92% (+ 4.9%). Average CAD sensitivity increases for solid, part-solid, and nonsolid NCNs were 5.1%, 4.2%, and 3.2%, respectively. Individual reader sensitivity improvements ranged from 1.5% to 14%. Seven of the nine readers had statistically significant sensitivity improvements at the .05 level. The CAD algorithm generated 5.3 false positives per case.

A prototype model-based CAD algorithm showed the potential to improve radiologist sensitivity for all NCN types by an average of 4.9% and as much as 14%.

E.A.K.,C.I.H.,J.K.L.: Claudia Henschke, Ella Kazerooni, and Joseph Leader participated in a study funded by General ElectricC.C.M.,W.T.,R.A.: Colin McCulloch, Wesley Turner, and Ricardo Avila are employees of General Electric

McCulloch, C, Henschke, C, Kazerooni, E, Leader, J, Turner, W, Avila, R, CAD-assisted Radiologist Sensitivity Improvement in the Detection of Solid and Subsolid Nodules in Low-dose CT Lung Cancer Screening Exams.  Radiological Society of North America 2004 Scientific Assembly and Annual Meeting, November 28 - December 3, 2004 ,Chicago IL. http://archive.rsna.org/2004/4405026.html