Binsheng Zhao DSc, Presenter: Nothing to Disclose

Yongqiang Tan PhD, Abstract Co-Author: Nothing to Disclose

Mingshi Wang, Abstract Co-Author: Nothing to Disclose

Hyun-Ju Lee MD, PhD, Abstract Co-Author: Nothing to Disclose

Chuanmiao Xie, Abstract Co-Author: Nothing to Disclose

Jing Qi, Abstract Co-Author: Nothing to Disclose

Ross C. Ehmke BA, Abstract Co-Author: Nothing to Disclose

Lawrence H. Schwartz MD, Abstract Co-Author: Nothing to Disclose

Radiogenomics promises the genetic assessment of cancer patients with non-invasive radiographic imaging studies. To date, little attention has been paid to the sensitivity of imaging features to repeat scans and acquisition parameters. This study explored the reproducibility of imaging features computed on repeat CT scans reconstructed at different parameters.

This study included a retrospective dataset containing 32 lung cancer patients, each having two same-day repeat CT scans and reconstructed into 6 image series, i.e., a combination of 3 slice intervals (5, 2.5 and 1.25mm) and 2 reconstruction algorithms [Lung (L) and Standard (S)]. Three radiologists independently used an in-house algorithm to segment 32 tumors (≥ 1 cm; one per patient) in all image series. 261 imaging features describing tumor size, histogram, shape, edge and texture were computed from the final tumor volumes, based on the common volumes obtained by 2 out of the 3 radiologists. The concordance correlation coefficient (CCC) was used to measure the agreement between each feature computed from two repeat scans reconstructed at the 6 series, i.e., 1.25L (first scan) &1.25L (second scan), 1.25S&1.25S, 2.5L&2.5L, 2.5S&2.5S, 5L&5L and 5S&5S.

Out of the 261 features, CCC of 32 features were ≥ 0.95 and of 169 features were ≥ 0.75 for all 6 series. Size and histogram features were highly reproducible for all parameter settings; shape index and boundary gradient strength were the least reproducible. For 2.5S&2.5S re-scans, all run-length, GTDM and spatial correlation features had CCC≥0.95; surprisingly, many features showed inferior reproducibility with 1.25L&1.25L re-scans, possibly due to noise.

Imaging parameters and repeat scans affect the reproducibility of imaging features to various degrees. Generally, the reproducibility of size, histogram, GTDM, run-length, spatial correlation, Laws', Gabor, wavelet, LoG and GLCM features (except 1.25L&1.25L for several features) at all parameter combinations ranged from acceptable (CCC≥0.75) to excellent (CCC≥0.95). The correlation of these features to gene expression warrants further investigation.

Precautions should be taken with regard to CT imaging acquisition parameters when conducting radiogenomics studies.

Zhao, B, Tan, Y, Wang, M, Lee, H, Xie, C, Qi, J, Ehmke, R, Schwartz, L, Reproducibility of Imaging Features Computed from Same-day Repeat CT Scan Images Reconstructed at Different Acquisition Parameters.  Radiological Society of North America 2013 Scientific Assembly and Annual Meeting, December 1 - December 6, 2013 ,Chicago IL. http://archive.rsna.org/2013/13024376.html