Understanding Uncertainties in Quantitative Image Features Extracted From Contrast-Enhanced CT Images
L Court*, J Yang, L Hunter, C Ng, MD Anderson Cancer Center, Houston, TXSU-E-I-58 Sunday 3:00PM - 6:00PM Room: Exhibit Hall
Purpose: Quantitative image features have been shown to be predictive of treatment outcomes for some patients groups. We examine two sources of uncertainty in these features for contrast-enhanced CT images of lung cancer patients: (1) The time between contrast injection and CT scan (which depends on local practice) (2) the reproducibility of features between scans.
Methods: Eight patients received two contrast-CT scans, 2-4 days apart. Each session included 6 repeat scans of the same anatomy taken every 20seconds, starting 50seconds after contrast injection. The timing of the scans encompasses the range expected during routine acquisition of contrast-CT images. For each patient, the lung tumor was manually delineated on the first scan, and propagated to each subsequent scan using deformable image registration. Image features based on tumor geometry, intensity histogram, absolute gradient image (IGR), co-occurrence matrix (COM) and run-length matrix (RLM) were then extracted from the tumor for each image set. Pearsons correlation coefficient was used to examine correlation between the features and time from contrast injection (each imaging session considered separately), with values over 0.60 considered to have a time-trend. The concordance correlation coefficient (CCC) was calculated to examine the reproducibility of each feature between sessions, with values greater than 0.90 considered reproducible.
Results: Variations in the image features attributed to cardiac pulsing was larger than any monotonic trends with time, and none of the image features were found to be correlated with time from injection. All geometric features and most histogram-based features (except maximum and minimum values) were found to be reproducible between scans. The majority of IGR-, COM- and RLM-based features were reproducible, and CCC was greater than 0.80 for all features.
Conclusion: The quantitative image features tested here were mostly reproducible, and have little dependence on the exact time between the contrast injection and the CT scan.