Medical Image Analysis Using Deep Neural Networks
With the improvement of clinical diagnostic equipment in hospitals and research institutions, large volumes of medical imaging data have been generated daily. However, the volume, variety and rate of generation of these images make it impractical and infeasible for clinicians to analyze them all without making any subjective errors. Therefore, the aims of this thesis are to enable computers to understand medical imaging data in the way doctors can and to help radiologists and physicians better use medical imaging data for automatic medical imaging diagnosis and intelligent healthcare for the benefit of clinics and patients.
To this end, the results of this thesis are expected to improve applications in medical image analysis fields as diverse as medical image retrieval, medical image and modality classification, medical tumor localization and segmentation, and tumor boundary delineation. Moreover, this thesis research seeks to address the major challenges in medical imaging data and gaps in current computerized medical image processing technologies, and then propose innovative solutions for automatic and accurate medical image analysis. The outcomes of the thesis will greatly enhance the smart use of medical imaging information and improve the ability of doctors' decisionmaking based on efficiently and precisely reading and interpreting the medical imaging data from various medical imaging modalities and generated scans. In addition, the outcomes will benefit to manage and access big medical imaging data and automate the process of medical image analysis with high accuracy and real-time response time for numerous clinical applications.
In particular, this thesis focuses on three related research areas for automatic medical image analysis, including medical image retrieval, medical image classification and medical image segmentation. The key contributions of the thesis are briefly summarized as follows:
· By leveraging preference learning and deep learning, a novel medical image retrieval system is proposed to accurately capture high-level semantic features of medical images and then efficiently index the similarly referenced images according to a given query image. Furthermore, this technique can also provide a key support to efficiently manage and store the newly generated medical imaging data for the development of medical imaging applications.
• In terms of limited medical imaging data, we have given an efficient and practical solution to address this challenge by utilizing transfer learning and data augmentation techniques. Furthermore, to cope with intra-class variation and inter-class similarity of medical images and modalities, a new multi-level featurefused convolutional neural network architecture is proposed to distinguish these imaging data.
• To address challenges from label noise and category imbalance problems of medical data to tumor feature selection, context information aggregation, evaluation of predicted outputs, we have devised a series of specific and targeted solutions, including a coarse-to-fine boundary delineation framework, a tumor attention network, a multi-scale fine-grained contextual information extraction architecture and a unified CTumorGAN framework.
• For going beyond the concept of current convolutional neural networks (CNNs}, we have proposed novel group equivariant segmentation CNNs by encoding more inherent symmetries existing in medical images. The work reveals a common bottleneck of the current segmentation networks for medical image segmentation.