Capsnet and ensemble of deep learning for medical image segmentation

Abstract

Medical images, such as X-Ray, Computed Topographic (CT) or Magnetic Resonance Imaging (MRI), requires expertise and patience to interpret. In clinical practice, only extensively trained specialists, also called radiologists, are qualified to read and provide observation report to physicians to assist their diagnosis. Due to the rapidly increasing demand of medical imaging, the radiologists are to handle a significant number of images every day. Thus, this is a huge pressure for them and life-threatening errors are likely to be made due to observers’ fatigue and rushing actions. Automatic medical image segmentation has been drawing considerable attention recently. This is because it has the potential to drastically reduce the workload of radiologists, automate and accelerate the process as well as improve diagnosis accuracy. Recent advances in deep learning techniques may make this technology possible to satisfy the strict requirements of clinical practice. Nonetheless, current state-of-the-art neural network methodologies still face a number of challenges. In particular, they lacks a sufficient amount of annotated medical data as qualified radiologists are often occupied. Errors and missing segments in the annotated data are also common. Medical images also tend to be too ambiguous for computer and data science professionals to understand without certain medical background. Finally, existing techniques to medical image segmentation also have inherent limitations.The objective of this project is to examine various deep learning approaches to medical image segmentation, investigate limitations in this line of research as well as propose novel methods to overcome current disadvantages and improve performances both experimentally and clinically. In the scope of the project, a variety of novel methods have been proposed. They were tested on three different segmentation tasks for medical anomalies and anatomical organs. Among them, four solutions have been experimentally shown to be effective and significantly con- tribute performance gain in segmentation. Two of them have been included into a conference paper entitled ”Medical Image Segmentation with Stochastic Aggregated Loss in a Unified U-Net”, which has been accepted for presentation at the 2019 IEEE EMBS International Conference on Biomedical & Health Informatics (BHI) (IEEE BHI 2019), Chicago, USA, May 2019.