Beyond the DRL: Applying Automated Quality Metrics to Assess Pediatric CT Program Liver Lesion Detection Performance


Conference Paper

PURPOSE To apply an automated program for evaluating pediatric body CT study quality which utilizes metrics of dose and image quality for optimization of liver lesion detection. METHOD AND MATERIALS With IRB approval, 880 clinical contrast-enhanced abdominopelvic (AP) CT scans of patients 0-18 years were evaluated. Studies were from Siemens Flash (n=621), GE 750 HD (n=151), and GE VCT (n=108). A quantitative metric of the detection of a potential 5 mm liver lesion was used as a marker for image quality (IQ). To generate this, metrics of spatial resolution, background noise, and lesion contrast were composited. Resolution was assessed by a validated method based on anatomical edges. For noise, phantom noise power spectra were matched to patient-specific scan parameters. For contrast, a 50 Hounsfield unit difference between the IV enhanced liver and a potential 5 mm lesion was the clinical task. The three quality metrics were used to calculate a single established detectability index (d’) which represents the relative likelihood of detecting the lesion and was previously correlated with observer performance. Dose reports were extracted for each dataset using an institutional dose monitoring program. Relationships between d’ and radiation dose were explored. RESULTS There was little CTDIvol variability across ages. For example, AP studies at 100 kVp on one scanner model had a median CTDIvol of 3.0 mGy (2.8-3.4 mGy interquartile range). However, when applying d', the age groups separated such that the younger patients had higher IQ than the older patients (Figure). For the youngest age group, d' and CTDIvol (medians) were 80 and 2.7 mGy; middle groups, 59 and 2.9 mGy; and oldest group, 42 and 3.4 mGy. CONCLUSION An automated method to assess clinical IQ using a task-based and patient-specific metric was ascertained. The d’ allows establishment of quality reference levels (QRLs) which account for IQ and dose. This provides for robust quality quantification that can serve for single or collective patient CT performance assessment and optimization. Automation also facilitates potential integration with CT registries and investigations using machine learning approaches not feasible with observer ratings alone. CLINICAL RELEVANCE/APPLICATION Optimization in CT utilizes DRLs based on dose estimates without metrics of image quality. The addition of quality measures taken from clinical examinations affords improved CT performance assessment.

Duke Authors

Cited Authors

  • Lacy, T; Ding, A; Abadi, E; Zhang, Y; Ria, F; Samei, E; Frush, DP

Published Date

  • November 25, 2018

Conference Name

  • RSNA 104th Scientific Assembly and Annual Meeting

Conference Location

  • Chicago (IL)

Conference Start Date

  • November 25, 2018

Conference End Date

  • November 30, 2018