Lesion Segmentation in 3D FDG-PET/CT Scans Using Deep Learning

Abstract: Drug development is an expensive, long and complex process for pharmaceutical companies all around the world. One way to make it more reliable is to evaluate the drug at its early stages, so the decision can be made to go forward or move to another drug, which is also a time-consuming and costly task to be done by human experts. Many studies have been done to automate this task since the machine is faster and less prone to error with repetitive tasks. A novel way in cancer drug development is utilizing deep learning andmachine vision techniques to segment cancer cells from radiologyimages and produce quantitative results to support the expertdecision on whether they continue with the current drug or not. This master thesis aims to search, explore, examine, modify andcompare the state-of-the-art performance of several Neural Networksarchitectures such as U-Net, Residual U-Net, and and AutomatedMachine Learning (AutoML) on Three-DimensionalFluorodeoxyglucose Positron Emission Tomography and ComputedTomography (FDG-PET/CT) scans of whole-body (head to thigh) formetastatic breast cancer patients. The results of the experiments show that the U-Net shaped networkshave stable performance and generalized even with a relatively smalldataset size, given the main criteria for comparison is Dice Score withfast and short training time in general