(Thursday, 8/2/2018) 7:30 AM - 9:30 AM
Room: Davidson Ballroom B
Purpose: To develop CT protocols that minimize uncertainties in radiomics studies.
Methods: We designed a radiomics phantom with 6 materials with textures similar to non-small cell lung cancer (NSCLC) patients’ tumors (from over 50 candidate materials). We determined acquisition and reconstruction parameters that minimized radiomics feature differences across manufacturers. Then, the phantom was scanned on 100 CT scanners (36 institutions) using this harmonized protocol and the local institution protocols for chest and head imaging. Radiomics features were extracted from the 300 scans semi-automatically. To assess variability, a linear mixed-effects model was then created with random effects for the phantom materials, and scanner nested within the manufacturer. The variabilities found with each protocol, and their relative components, were compared. Additionally, the variability in features resulting from tube voltage and current changes for tissue-equivalent materials was compared to the interpatient variation found in 20 NSCLC patients.
Results: Image thickness and pixel size varied greatly across local institution protocols: median 3mm (range:0.5-5mm) and 0.7mm (range:0.3-1.6mm), respectively; other parameters also varied widely. Image thickness was correlated with feature values and had to be included as a fixed effect in the linear mixed effects model (average absolute Pearson correlation 0.42, range:0.02-0.8). Resampling the image prior to feature extraction did not meaningfully alter the feature values (ICC>0.9). The harmonized protocol greatly reduced variability with 57% and 52% less variability than the local chest and head protocols, respectively. Furthermore, 74% and 88% of features were not impacted by changes in tube voltage and tube current, respectively.
Conclusion: A harmonized protocol can reduce interscanner variability by over 50%, and we recommend its use in prospective radiomics studies. As many acquisition parameters do not impact features, the majority of these improvements can be achieved by simply including an additional reconstruction with existing imaging protocols with no extra patient dose.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by the NCI Grant R21CA216572.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
