Applications of Reinforcement Learning to Medical Image Segmentation


This project investigates the application of reinforcement learning techniques to the segmentation of medical images. In particular, we present a novel approach that is based on learning how to grow a selection with regions obtained from an image partition forest (IPF) based on various attributes of the regions. Our algorithm will be almost automatic, although we discuss why it cannot be classed as entirely automatic. We then proceed to quantitatively evaluate this method against two datasets of manually segmented femurs in MRI scans of knees and compare its performance to that of state-of-the-art algorithms. We also discuss advantages that our approach provides, such as achieving results in a much shorter time and with significantly less expensive data labelling.

Masters Project Report