Quantification of hepatic steatosis using machine learning and color image processing

Hepatic steatosis -the accumulation of fat in the liver- is a risk factor for poor function liver transplantation and the main cause of donor rejection.

The assessment of steatosis is a difficult task for the transplant team. An initial evaluation, based on visual inspection is first done during the liver procurement. However, criteria such as color and texture depend on the experience of the surgeon, and thus remain subjective and prone to errors. These errors are problematic in two aspects: First, when transplanting a liver with hepatic steatosis the donor rejection rate escalates. Second, due to the lack of donors, only around 50% of people in the waiting list will be able to receive a liver. Therefore, rejecting livers that are viable reduces the live expectancies for people in the waiting list. The only current fully accurate procedure for diagnosis and staging is liver biopsy, but is invasive and costly.

This project aims to develop tools to automatically detect and quantify hepatic steatosis to increase the pool of valid hepatic donors by objectively and quantitatively evaluating the degree of hepatic fat. To this end, we will analyse different color representations (such as CieLab, CIECAM00, HSV) together with different contrast enhancement techniques and texture descriptors. All these priors will be combined with different machine learning techniques, ranging from straighforward Support Vector Machines to more complex optimizations such as Convolutional Neural Networks (CNNs) in order to obtain the best possible accuracy in measuring the hepatic steatosis.