TISSUE CLASSIFICATION AS NON- INVASIVE ASSESSMENT METHODS FOR LIVER STEATOSIS

Hepatic steatosis (HS) is one of the most important organ donor characteristics that can influence negatively graft function and so liver transplantation (LT) outcome. Despite histopathological analysis of liver tissue is currently the gold reference standard for diagnosis and grading of HS in liver grafts, it is invasive, time-consuming and expensive. Due to the short time availability between liver procurement and transplantation, the surgeon usually performs HS assessment through clinical evaluation (medical history, blood tests) and qualitative visual liver graft assessment. In this context, visual liver texture analysis is recognized as crucial in grading HS: livers that cannot be transplanted due to high HS are usually characterized by inhomogeneous texture and are more yellowish than the transplantable ones. It is nonetheless recognized that the precise estimation of HS remains challenging even in experienced surgeon’s hands. The automatic image-based intraoperative tissue classification is a valuable solution to provide decision support to the liver surgeon and context awareness intra-operatively. The surgical data science (SDS) community is focusing more and more on machine learning (ML) to perform tissue classification in the operating room (OR). However, several technical challenges as noise in the image, varying illumination level, different camera pose with respect to the tissues, and intra- and inter-patient tissue variability, are still present, hampering the clinical translation of the developed methodology in the clinical practice. On this background, the overall goal of this PhD thesis was to tacke most of theese issues by developping an hardware device able to allow high-quality standard intraoperative image acquisition and a framework for ML-based SDS algorithms for robust and reliable tissue classification for intraoperative images. In particular, the contributions of this PhD work are: 1. A new tool for intraoperative RGB images acquisition 2. A new method for reliable hepatic tissue classification that utilised images acquired with the new device and non-invasive biomarkers donor’s data 3. A machine-learning based evaluation of non-invasive biomarkers donor’s data with the goal of estimate the real value of clinical and biochemical donor’s features in HS classification 4. A new method for hepatic tissue classification arising from human visual cortex method, that utilised only images acquired with the new device designed This PhD thesis work lies in the wider research field of SDS, which aims at providing the surgeons with decision support and context awareness during interventional medicine procedures. The methodologies presented in this PhD thesis demonstrated the feasibility of using SDS ML-based algorithms for intra-operative hepatic tissue classification in liver transplantation setting. The methodological progress made in this work highlights the potential of SDS ML-based algorithms in extracting useful information implicitly contained in intra-operative optical images, overcoming challenges typical of intra-operative tissue classification to support surgeons during liver transplanatation procedures.