Sparse Coding for Medical Image Analysis: Applications to Image Segmentation and Classification

Abstract

Medical imaging is a fundamental and invaluable tool in modern healthcare. The use of medical imaging has greatly increased, and these massive image archives provide the opportunities for evidence-based diagnosis, physician training and biomedical research. The ability to automatically analyse, quantify, and categorise the images within these data archives is an important enabling technology for computer-aided diagnosis (CAD) system which can be used to aid in patient care. For example, automated lesion segmentation and disease classification of medical images have shown greater clinical benefits and provided second opinion to diagnosis. Such systems are predicated on the ability to relate low-level image features to high-level semantic concepts or expert domain knowledge. These systems are dependent on the use of pre-defined low-level image features such as colour, texture, and shape, to derive high-level semantic concepts (e.g. expert domain knowledge). However, developing robust and accurate CAD systems is challenging as the analysis and diagnosis are often made using the semantic features that cannot be effectively represented by the pre-defined low-level features only. Representing medical image in a semantic space that captures the essence of the image is an important and open challenge, and involves finding a most appropriate image representation and matching techniques for each CAD system. In recent years, sparse coding has shown a great effectiveness in understanding and capturing the rich semantic information within natural images and signals. Sparse coding provides a class of learnable algorithm that captures higher-level features from unlabelled input data; the learned algorithms (e.g. bases function) resemble the receptive fields of neurons in the visual cortex. As such, in this research, we investigate sparse coding algorithms within medical images to effectively capture salient and semantic information of medical images. In addition to sparse coding algorithm, we further investigate saliency detection algorithms; they are motivated by how human visual system works and aim to detect informative, interesting, and salient areas in an image. This helps to find lesions or tumours that help to understand and represent an image. We also explore convolutional neural networks (CNNs) which uses multiple deep layers of non-linear information for pattern analysis and classification. CNNs learn image features in a hierarchical fashion; each deeper layer of the network learns a semantically “higher” level representation of image data. Specifically, this thesis focuses on the application of medical image segmentation and classification using different medical images of Dermoscopic images and X-ray images. We show that salient and semantic features in these imaging modalities can be captured via sparse coding algorithm together with a set of new image processing algorithms such as saliency detection and CNNs. In this thesis, we firstly proposed a novel algorithm for the segmentation of skin lesion in Dermoscopic images. Our new sparsity and saliency based algorithm exploits the reconstruction errors derived from sparse coding coupled with a novel background detection optimized for Dermoscopic images. We also propose a new classification algorithm for the categorization of X-ray images. Our method uses a late-fusion of sparse spatial pyramid (extension of sparse coding) with domain transferred convolutional neural networks (DT-CNNs). Our method is robust as it exploits specific local features and sparse characteristics inherent in the X-ray images, and uses the rich generic information provided by the DT-CNNs. We evaluated our proposed algorithms on public datasets and compared them to other common state-of-the-art methods. Our results demonstrate that our algorithms were more accurate and robust in both segmentation and classification compared with other methods.