Evaluation of diagnostic accuracy: multidetector CT image noise correction improves specificity of a Gaussian model-based algorithm used for characterization of incidental adrenal nodules

Abstract

Objectives

To investigate whether the histogram analysis method of characterizing adrenal nodules as adenomas is affected by increased noise with modern CT technique, and if an extension that allows for noise correction will improve diagnostic performance.

Materials and methods

This is a HIPAA-compliant, IRB-approved retrospective study performed on 58 total patients. The first group of 29 patients had 33 adrenal lesions that were pathology-proven non-adenomas. The second group had 29 patients with 33 pathology-proven or presumed adenomas based on established imaging criteria. The nodules were evaluated using the histogram method, mean attenuation method, and a Gaussian model-based algorithm without (uncorrected Gaussian algorithm) and with correction (corrected Gaussian algorithm) for image noise. Sensitivity, specificity, and accuracy for identifying adenoma were derived.

Results

There were no significant differences in identifying adenoma from non-adenoma when using the histogram analysis method and the uncorrected Gaussian algorithm, both of which had low specificities of 42.4% and 47.0%, respectively (p = 0.30). Adding noise correction to the Gaussian algorithm resulted in a statistically significant increase in specificity relative to the histogram method (86.4% vs. 42.4%, p < 0.001). The corrected Gaussian algorithm improved sensitivity compared to the mean attenuation method (71.2% vs. 54.5%, p < 0.001), but had lower specificity (86.4% vs. 100%, p < 0.001), and similar overall accuracy (78.8% vs. 77.3%, p = 0.74).

Conclusion

With modern low-dose CT technique, the specificity scores of the histogram method for discrimination of adrenal adenomas and non-adenomas are lower than with previous higher dose scans. The specificity and accuracy of a histogram-equivalent method can be increased mathematically through image noise correction, and the corrected Gaussian algorithm has improved sensitivity to the mean attenuation with similar accuracy albeit with lower specificity. Although this suggests limited utility for histogram analysis in adrenal nodule characterization, our study demonstrates the potential mathematical application for other noise-dependent CT characterization methods.