Click here to


Are you sure ?

Yes, do it No, cancel

Step-Wise Solution to Evaluate CT Radiomic Feature Variability Due to Correlated Noise Texture

M Shafiq ul Hassan1*, F Guo1 , H Chen1 , G Zhang2 , E Moros2 , Z Chen1 , (1) Yale New Haven Hospital, New Haven, CT, (2) Moffitt Cancer Center, Tampa, FL,


(Sunday, 7/14/2019) 4:30 PM - 5:00 PM

Room: Exhibit Hall | Forum 9

Purpose: Radiomics is an emerging field with promising prospects for detection and therapy response assessment in oncology. One problem in radiomics is the feature variability due to correlated noise in reconstructed CT image. The variability of texture features has been reported due to reconstruction kernel. As most texture features describe spatial relationship between neighboring voxels, it is possible that texture features are also sensitive to correlated noise due to reconstructed slice thickness. Similarly, noise texture due to pixel size variation could be another source of variability in second and higher order texture features. Here, the purpose is to highlight the problem of correlated noise in CT radiomics and present its solution.

Methods: This problem can be solved using a step-wise approach: (a) calculation of feature variability due to each parameter using a texture phantom; 162 scans of Credence Cartridge Radiomic Phantom with varying pixel size and slice thickness are publicly available on TCIA website (b) quantification of noise texture by measuring NPS using an ACR phantom (Model-464); third module having uniformity portion can be used for this purpose and (c) correlate feature variability with Noise power spectrum (NPS). In this work, 69 features were extracted using an in house program. Feature variability with correlated noise was evaluated using concordance correlation coefficient.

Results: Most higher order features (NGTDM, GLSZM) were found to be correlated with correlated noise, produced by reconstructed slice thickness and pixel size. However, most first order and second features were found to be robust with respect to correlated noise.

Conclusion: Correlated noise texture can be quantified using the noise power spectrum. Feature variability due to correlated noise can be assessed using method described here. In future, this method will be expanded to evaluate the correlation of 1100 features with correlated noise.


Data Acquisition, Quantitative Imaging, Reconstruction


IM- CT: Phantoms - physical

Contact Email