UPmémoires

English

Objectives: To investigative the accuracy of radiomics features to cholegallstone depiction from pre-contrast computed tomography (CT).

Background and objectives: Standard abdominal CT is the first step imaging modality for acute abdominal pain. It is not sensitive for cholegallstones depiction. From the seventies, texture analysis has tried to provide a quantitative assessment of density heterogeneity by analyzing the distribution and relationship of pixel or voxel gray levels in the image. In liver imaging, there are numerous potential applications of this technique, often called “radiomics”, such as tumor grading, fibrosis assessment or evaluation of tumor response. However, to date, there is no available data on the potential interest of pre-contrast CT radiomics analysis in biliary gallstones depiction. Our hypothesis is that non radio-opaque lithiasis might provide density changes only depicted with radiomics analysis. We would like to evaluated the accuracy of radiomics features in cholegallstone depiction from pre-contrast computed tomography. MRI was used as gold standard method.

Methods: We included patients admitted to our University Medical Center between March 2014 and March 2017 who had abdominal CT and MRI within 45 days. Twenty-two patients were excluded due to a small gallbladder (less than 1cm in axial diameter - threshold for our in house software). A senior radiologist with 10 years experiment reviewed T2W sequences of abdominal MRIs. Pre-contrast CT images were blindly and independently reviewed by a 7 years experiment radiologist and a junior fellow. For the discordants cases, a third review was realised by an independent radiologist. Furthermore, an inter observer agreement of pre-contrast CT reports was assessed using a kappa test. Blindly to MRI, the gallbladders contents were segmented with the software 3DSlicer® for each patient and a radiomics analysis was performed on the volume of interest by an in-house software. 810 radiomics features were generated for each patient. The ability of the radiomics features to differentiate patients with or without gallstone were established using receiver-operating characteristics (ROC) curve. Area under curve (AUC) has been calculated. A univariate and a multivariate analyses were performed. Association between radiomics features and gallstone were explored using all cohort as well as separating cohort into training (60%) and testing (40%). Receiver operating characteristics (ROC) curve was used to evaluated accuracy to the radiomics features.

Results: Using MRI analysis, 39 patients had cholelithiasis and 61 not. Using visual CT analysis, 24 patients had cholelithiasis and 76 not. The sensitivity, specificity and AUC were respectively 56.4% [39.6 – 72.2; CI 95%], 96.7% [88.6 – 99.6; CI 95%] and 0.77 [0.67 – 0.84; CI 95%]. Using pre-contrast CT-derived radiomics features, in univariate analysis, the highest AUC of 0.8 were obtained. In multivariate analysis, AUC of 0.87 were obtained using all dataset. Whereas, in multivariate analysis, in training AUC of 0.87, sensitivity 93% and specificity 80% were obtained and AUC of 0.76 in validation. Radiomics features were independent predictor for cholegallstone depiction, which could successfully categorize patients with or without gallstone. They improved the performance compared to classical visual assessment of pre-contrast CT.

Conclusion: Radiomics features seems to help the detection of gallbladder gallstone using pre-contrast CT compared to classical visual assessment of CT.