Efficient deep learning based assisted annotation for medical image segmentation

Medical image segmentation is widely recognized as one of the most challenging tasks in medical image analysis. It involves identifying the set of voxels that constitute organs or lesions of interest from the background. Accurate segmentation is crucial for conducting quantitative measurements and analysis of clinical parameters such as volume, shape and intensity distribution. Furthermore, it enables physicians to monitor disease status and progression, predict disease course and select the most appropriate treatment plans in clinical practice. To improve segmentation accuracy and reduce the burden on experts, various semi-automated and automated approaches have been proposed. Recently, deep learning methods based on artificial neural networks have revolutionized medical image analysis and achieved state-of-the-art results in many challenges. However, several impediments remain. Firstly, deep learning results may be imperfect on explicit topology and require further improvement. Secondly, accurate annotation of medical image datasets is pivotal in deep learning, but the process is not only tedious and time-consuming but also demands costly, specialty-oriented knowledge and skills that are not easily accessible. Therefore, this dissertation aims to address these critical problems by developing efficient deep learning algorithms for medical image segmentation when annotated datasets are limited/unavailable and effectively assisting human experts to decrease annotation effort. This dissertation outlines three specific aims from different perspectives: (1) Develop novel methods for efficient assisted annotation in the Deep-LOGISMOS-JEI framework to utilize existing annotation effectively for improved segmentation; (2) Develop novel strategies to utilize automatic segmentation quality assessment (SQA) to acquire necessary annotation efficiently to reduce the demand for training data size and assist experts in annotating images effectively; (3) Develop novel self-supervised learning approaches for deriving representations/features from large unannotated medical image datasets to improve segmentation performance on small annotated datasets. The effectiveness of the proposed methods was demonstrated by experiments on calf muscle segmentation on magnetic resonance (MR) images, vessel wall and lumen segmentation on intravascular ultrasound (IVUS) images and psoas muscle segmentation on computed tomography (CT) images.