M.Sc Thesis


M.Sc StudentGoldreich Gilad
SubjectDetecting Severely Altered Biodistribution in FDG PET/CT
Scans using Machine Learning and Deep Learning
Techniques
DepartmentDepartment of Industrial Engineering and Management
Supervisors ASSOCIATE PROF. Tamir Hazan
DR. Michal Weiler Sagie


Abstract

18F-Fluorodeoxyglucose positron emission tomography (FDG-PET/CT) scans are the

current workhorse of nuclear medicine departments world-wide. A diagnostic PET component enables detecting the high uptake in pathological processes such as tumors and metastases, where the ability to detect abnormal uptake foci is based on high uptake in pathological cells relative to background uptake in normal organs. The background uptake is a testimonial of the glucose utilization of the different organs or excretion pathways of FDG, and results in the normal biodistribution of FDG. Unfortunately, altered biodistribution of tracer in the body may diminish PET quality as it results in increasing background uptake in normal organs, thus reducing conspicuousness of cancer tumor and metastases. It also raises concern for reduced availability and therefore reduced uptake in the cancer cells, further hampering the diagnostic capability of the PET scan, possibly yielding false-negative results. The detection of altered biodistribution in a PET scan is a qualitative, subjective impression based on the interpreting physician's knowledge of the normal biodistribution pattern, yet objective assessment of abnormal biodistribution and detection of severely altered biodistribution can improve the interpreting physician’s confidence in their report and in the important decision of repeating the scan. In this research, we’ve explored machine learning and deep learning techniques in order to generate a model which will attempt to predict, given an FDG-PET/CT scan, if the FDG muscular biodistribution is severely altered. Driven by the need to perform rapid processing of scans, we have created an efficient, lightweight PET/CT viewing and tagging software, designed to handle multiple scans at a time. The effectiveness of the proposed method was validated on a specially made PET/CT dataset of 2727 patients, collected during collaboration with Rambam Health Care Campus. The results demonstrated that deep learning models can detect severely altered biodistribution in FDG-PET/CT scans, which can even be expanded in order to perform inference on scans that are much harder to interpret without manual adjustment due to sub-optimal contrast level, thus enabling the model to perform predictions without any need of human intervention. Such a tool has the potential to improve patient care and to provide insight into patient specific and time specific glucose metabolism.