Optimizing CT-Based Models for Predicting Whole-Body Fat in Rabbits.

Abstract

This study evaluated three computed tomography (CT)-based automatic fat segmentation methods and identified the optimal regression model for predicting whole-body fat in rabbits. CT scans were performed on 21 postmortem companion rabbits to estimate fat percentage (Fat%-CT) and fat volume (FV-CT, mm³). Two techniques were used to determine CT number (Hounsfield units [HU]) ranges for fat (HU Ranges 1 and 2). These were applied to four body regions using three voxel counting techniques (one-, two-, and three-dimensional techniques) to quantify Fat%-CT and FV-CT. Data from whole-body chemical carcass analysis (whole-body Fat% and whole-body fat in grams) were used as gold standard values. The relationship between the extracted CT data and carcass fat estimates was determined using a Spearman rank correlation. Simple linear regression and assumption tests were performed to identify the best model for predicting whole-body Fat%, and whole-body fat (g). There was a high correlation between measured fat estimated by CT and by chemical carcass analysis (r_s > 0.7, p < 0.001) for all but one body region-counting combination. The best model predicting whole-body Fat% used CT data from left kidney region, HU Range 2, and three-dimensional counting technique. The optimal model predicting whole-body fat (g) used CT data from the entire abdominal region, HU Range 1, and a one-dimensional counting technique. The CT-based models for estimating whole-body fat in rabbits are feasible. Model optimization requires appropriate selection of the body region, CT number ranges, and counting techniques.