LIVERCOLOR: An Algorithm Quantification of Liver Graft Steatosis Using Machine Learning and Color Image Processing

Background: The use of liver donors with significant macrosteatosis is associated with a worse outcome and this is one of the major reasons to decline liver grafts for liver transplantation. Deceased liver donor acceptance is mostly based in subjective surgeon criterion of liver aspect, due to macrosteatosic livers acquire a yellowness feature. Aim: To develop a machine learning model using a color mobile calibrated images for the determination of macrosteatosis in liver grafts. Methods: The algorithm has been developed in a University Department for New Medical Technologies. For each deceased donor a total of 5 pictures were taken. The photography was carried out using mobile telephony systems with a 12-megapixel camera. The surgeon had to place a previously sterilized flat grey plastic checker card next to the liver to be photographed, oriented toward the direction of the incident light of the mobile camera and all the images were taken at a distance of 10 cm from the liver. All the grafts underwent two separate tru-cut needle biopsies, one for the left lobe and the other for the right one. Steatosis was assessed based on the percentage of hepatocytes with macrovesicular steatosis as mild (< 30%) or moderate-severe (≥ 30%). All liver images were color calibrated and segmented and a feature extraction based on L*a*b* color space (L* represents luminescence layer, a* chromatic layer for the red-green edge and b* chromatic layer for the yellow-blue edge) and Local Binary Pattern was performed. Results: Forty-two donors were included for training and testing cohorts and 73 liver images were retained. Eleven images were excluded due to slow-quality image resolution. Sixty-hundred and forty-five liver patches were performed (344 from the right liver and 301 from the left liver). The best learning model was obtained with the a* chromatic as predictor. It has an accuracy of 85,3% and the specificity for the moderate-severe grades is of 98%. Conclusions: LiverColor machine learning has an excellent accuracy and specificity to determine liver macrosteatosis. Image Legend: flowchart of the study