Our results suggest that lung cancer screening should be performed by a radiologist using any CAD.
Combined CT-sensitivity does not drop significantly at lower dose levels,
including 25mAs/80kVp with an average dose length product of 22.3 mGycm,
equaling an equivalent dose of 0.3 Millisievert.
In a previous study (under review),
we stated that a minimum dose level of 25 ref mAs/100kVp is needed to detect both solid and ground glass nodules.
This threshold level applies to GGNs when analyzed separately,
but in a lung cancer screening setting where the prevalence of GGNs is set to 25% [21-23] the voltage could be reduced to 80 kVp.
The individual performance of both radiologists and CADs did not drop significantly at lowest dose level,
except for CAD3.
The interspersed GGNs were significantly less detected by CAD,
especially by CAD3,
which probably strongly influenced the significance level.
Otherwise,
our results are consistent with previous studies stating that low dose CT is not significantly affecting the performance of CAD [5,
15,
31].
Sensitivity of radiologists is higher due to the better detectability of GGNs compared to CAD.
Best double-reading pair was a radiologist with a CAD; sensitivity was always over 96% at standard CT independent of the radiologist and the CAD,
when at least one of them had to detect the nodule.
This pairing was better than any combination of CADs or the pairing of two radiologists.
The radiologists - although scoring high individual sensitivities - detected the same nodules (excellent inter-observer agreement),
therefore the addition of a second radiologist was lower.
The CADs had lower sensitivities; therefore a combination of two CADs did not reach the combined sensitivity of a radiologist and a CAD.
Inter-observer agreement between radiologists was significantly higher than the agreement between radiologists and CADs,
suggesting that more additional information (sensitivity) would come from a CAD as a second reader than from a second radiologist,
as long as the sensitivity of the CAD is acceptable.
Although radiologists had better sensitivities for lung nodules,
the combined sensitivity rose only for one reader significantly when a second radiologist performed the double reading.
The benefit of double reading by radiologists (+4.5% and 6.5%) was not as high as described by Wormanns et al (11-15%) probably because of their lower baseline sensitivity of 63% [5],
which leaves more room for additional sensitivity.
When the CAD was the second reader,
the combined sensitivity was significantly higher than the individual sensitivity for both radiologists almost for all dose levels.
The low agreement between CAD and radiologists overcomes the slightly lower sensitivity of CAD to detect lung nodules.
The prototype CAD of Philips scored the lowest sensitivities and the lowest agreement with the radiologists and reached the same combined sensitivities as the other CADs.
We therefore confirm the additional value of a CAD stated by several authors [6,
9,
11,
12].
In our study the mean radiologists’ sensitivity gained using a CAD was 6.9% and 10.3% for standard and lowest dose examined.
This additional sensitivity is not as much as the 14% stated by Li et al [6],
but their baseline sensitivity was 52%,
compared to our baseline sensitivity of 82-88% at lowest dose and 90-92% at standard dose.
Furthermore,
the added GGNs with their lower detectability may have influenced the benefit for the worse.
Low inter-observer agreement and high sensitivities on both observer sides would be the optimal reader pairing and the ratio of sensitivity divided by the agreement (mean sensitivity of both sides/Kappa) could be an important indicator for additional nodule detectability of a CAD-system to radiologists.
Limitations:
We did not analyze the additional value of maximum intensity projection as second reader.
This is subject of ongoing studies at our institution.
At the time of our data acquisition there was only filtered back projection available for image reconstruction at our institution.
Meanwhile the iterative image reconstruction reduces radiation exposures in our daily routine: preliminary results indicate a dose reduction of 50% at the lower dose levels with the same image quality and noise; meaning a potential low dose CT feasibility below 10 ref mAs/80kVp.
Acknowledgement:
The Bernese Cancer League,
the Jubilee Foundation of SwissLife and the Swiss foundation “Fight against Cancer” financially supported this study.