A study of deep learning for automatic segmentation of healthy liver in abdominal computed tomography scans

Abstract

Medical image segmentation refers to a process in which Regions of Interest (ROI) such as organs are annotated in 2D or 3D medical images. Medical image interpretation performed by radiologists and physicians has proved to be crucial for early clinical detection, diagnosis, and treatment; however, precise manual segmentation of medical images is a time-consuming process, when shortening the time between medical scanning and any required medical procedure is critical. As a result, the use of computers to assist medical image interpretation, Computer-Aided Diagnosis (CAD), has become compulsory to assist radiologists. In recent years automatic image segmentation based on Deep Learning (DL) models have become popular due to the fast and precise results that can be achieved, surpassing tradition methods. Despite DL models surpassing traditional methods, the segmentation required in the medical industry requires precision that DL models do not yet reach. The high variability from patient to patient, ROI overlapping, limited size of datasets to learn from, and low-resolution images, are characteristics which make it difficult to develop a universal DL model suitable for any specific problem. In this project a review of existing solutions, architectures and implementations for medical image segmentation was first carried out, and key concepts were highlighted. State-of-the-art Deep Neural Networks were implemented, namely U-Net and 3D-UNet. A proposed model was then implemented, where medical images are first pre-processed and classified to include liver before passing through the Deep Neural Network. The models were applied and compared to the liver CT scan dataset publicly provided by the CHAOS challenge, where the CT scans were acquired at portal phase after contrast agent injection for pre-evaluation of living donated liver transplantation donors. The proposed model was found to improve over the precision of U-Net, from a DICE score of 0.94 to 0.97.