E2375. Identification of Radiomic Correlates of Hematology and Serum Chemistry markers of Ebola virus disease
  1. Bino Varghese; University of Southern California
  2. Jeffrey Solomon; Frederick National Laboratory for Cancer Research
  3. Steven Cen; University of Southern California
  4. Nina Aiosa; Center for Infectious Disease Imaging
  5. Syed Reza; Center for Infectious Disease Imaging
  6. Sean Bartlinski; Integrated Research Facility
  7. Vinay Duddalwar; University of Southern California
A non-invasive longitudinal assessment method for Ebola virus disease (EVD) status and prognosis are urgently needed. We identify strong radiomic correlates of hematology and serum chemistry markers of EVD using CT imaging data from infected nonhuman primates.

Materials and Methods:
In this pilot study, longitudinal CT images from 17 primates infected with Ebola virus (EBOV) were acquired. 3D contours of the liver were manually segmented using MIM 6.7.9 software (Cleveland, OH) and in a subset of cases automatically segmented with a deep learning-based convolutional neural network (feature pyramid network). The segmented regions were then used as inputs in the CAP-TK software (U Penn) to perform radiomics analysis. For each primate, 1960 radiomic, 30 serum chemistry and 47 hematology metrics were extracted at baseline and at least one follow-up time point when imaging was available. Time dependent intra-subject correlation analysis was performed to identify the family of radiomic metrics with the strongest correlation with changes in hematology and serum chemistry panel metrics of EVD. A correlation Heatmap was used to visualize relationships with strong correlation coefficient (0.5 <= r <= 1 and -0.5 >= r >= -1).

Of the various radiomic metrics considered, Grey-level run-length method (GLRLM) (39%), Histogram analysis (26%) and morphologic analysis (15%) showed the highest number of strong correlations with the hematology panel in EVD. GLRLM metric: LowGreyLevelRunEmphasis_Offset_11_Kurtosis was highly negatively correlated (r>-0.8) with albumin/globulin ratio. Strong correlations (r<-0.66) were also found between 1st order texture metrics such as histogram-based interquartile range, variance, 95_Percentile_Median and 95_Percentile_Mean and, liver enzymes such as aspartate aminotransferase (ALT), alanine aminotransferase (AST), gamma-glutamyl transpeptidase (GGT) and total bilirubin, respectively.

The correlation of radiomic features of the liver, often damaged during EVD, with albumin and globulin is relevant to EVD as albumin is made by the liver and keeps fluid from leaking out of blood vessels while globulin, made by liver and immune system help fight infection. In addition, the radiomic correlation with liver function tests such as ALT, AST, GGT and bilirubin in EVD is encouraging. Radiomic correlates of hematology and serum chemistry panel metrics demonstrate the potential for using radiomics from CT scans as a surrogate marker for EVD outcome.