A long-standing goal of the medical community is to present and analyze medical images efficiently and intelligently. On the one hand, it means to find efficient ways to acquire high-quality medical images that can readily be used by healthcare providers. On the other hand, it means to discover intelligent ways to interpret medical images to facilitate the healthcare delivery. To this end, researchers and medical professionals usually seek to use computerized systems that are empowered by machine learning techniques for the processing of medical images. A pivotal step in applying machine learning is to obtain informative representations that well describe medical images. Conventionally, this is performed with manual feature engineering which however requires considerable domain expertise in medicine. A possible workaround is to allow the model to automatically discover latent representations about the target domain from raw data. Recent advances in deep learning have made unprecedented progress in this direction. Deep learning exhibits incredible ability in discovering intricate structures from high dimensional data. For many computer vision tasks, deep learning approaches have achieved state-of-the-art performance by a significant margin.
This thesis will develop deep learning models and techniques for the interpretation, reconstruction, and generation of medical images. In medical image interpretation, we concentrate on understanding the content of the medical images and giving guidance to medical practitioners. In particular, we investigate deep learning ways to boost the diagnostic classification, landmark detection and substructure segmentation from medical images. In medical image reconstruction and generation, we propose to use deep generative models to inherently learn the medical data space and effectively synthesize realistic medical images. For the reconstruction, we aim to generate high-quality medical images with fewer artifacts and without sacrificing the acquisition time. For the generation, our goal is to generate realistic medical images to help the learning of medical image interpretation models.
In this proposal, we present our preliminary works that leverage deep learning to solve various problems in medicine. Specifically, we demonstrate 1) how deep neural networks can be used to better interpret medical images for dermatology diagnosis, vertebrae localization, and cardiac structure contouring; and 2) how deep generative models can be applied for cone-beam computed tomography artifact reduction. In our future work, we plan to develop deep learning models to cover a broader range of medical problems. Namely, we expect to explore deep learning solutions to pseudoprogressive tumor identification, metal artifact reduction, and multimodal magnetic resonance imaging modality transformation. Meanwhile, we will keep track of other possibilities in developing and applying deep learning in medicine. Timelines for the future thesis work is included at the end of this proposal.