Architectural and algorithmic design for embedded medical imaging

• Main Author: Chiew, Wei Ming
• Other Authors: Lin Feng
• Format: Theses and Dissertations
• Language: English
• Published: 2014
• Subjects: DRNTU::Engineering::Computer science and engineering::Computer applications::Life and medical sciences; DRNTU::Engineering::Computer science and engineering::Computing methodologies
• Online Access: https://hdl.handle.net/10356/61735
• Institution: Nanyang Technological University

We are motivated by the lack of a coupling mechanism between imaging, registration and rendering for an efficient process cycle, which is critical to realize 3D structures and depth-correspondences across slices for real-time quality assessment and diagnosis. Temporal slice capture mechanisms of imaging platforms also present a unique standpoint which differs from the operating framework of offline registration and rendering methods. Literature survey revealed a lack of research effort in a real-time registration-rendering solution alongside acquisition. Moreover, solutions for non-rigid registration with embedded hardware are almost non-existent. Therefore, this thesis addresses these problems by integrating the three pillars of modern confocal imaging, including image acquisition, image registration and volume rendering into a single FPGA architecture for customizable prototyping. The main objective is to provide a novel framework for real-time registration-rendering alongside acquisition which is targeted ultimately at application-specific integrated circuit design. A pioneering research project collaborated with the National Cancer Centre Singapore to develop an embedded rendering solution for confocal endomicroscopic imaging justified the need for rendering of registered datasets captured in vivo. Motion exhibited due to the nature of living tissue distorts the volumetric dataset and impedes accurate visualization. Consequently, we propose a processing framework targeted at embedded architectures and detail functional requirements and desired specifications of the system. The study is conducted and verified with FPGA prototyping platforms. Next, a slice-based registration approach is presented with two different modes of operation: inter-dataset and intra-dataset registration. For embedded processing, a first-of-a-kind non-rigid registration algorithm on FPGA is demonstrated and accelerated using a novel single-pass Demons kernel design and tightly-coupled regularization. A 14-fold reduction in computations and 175 times speedup are achieved with each method respectively. A novel scheme for real-time rendering called online incrementally accumulated volume rendering has been proposed. To manage data transfer between acquisition and rendering, an automated mechanism for image capture is described, alongside optimization methods including render space consolidation, cubic memory organization and accumulative rendering mode. An experimental implementation shows real-time performance of 50 frames per second with maximized utilization and horizontal scaling. Finally, we propose an interpolation method to couple registration with rendering. A novel interpolation method called negative square distance is detailed. Two configurations for deformable interpolation are described for object-ordered and image-ordered rendering respectively. Quantitative experimental analyses of resource utilization, timing and throughput, and scalability potentials are presented with rendering results of the methods. This PhD study makes the following contributions: (a) the novel framework for on-the-fly registration-rendering with slice-volume image acquisition systems provides new insight to both the clinical and biomedical community as well as to computer vision and visualization disciplinarians; (b) although research is mainly done with Demons registration, the improvement methods can benefit iterative registration algorithms in general as well; and (c) the novel interpolation algorithm developed may have a wider audience especially for real-time applications with embedded systems such as computer-aided surgical systems, computer vision systems and interactive ubiquitous systems. Clinical dataset examples in respective chapters show the value of our approach in practice.