Research On Iterative 3D Tomography Reconstruction And Feature Detection Based On Approximate Computing

3D tomographic reconstruction, which requires a three-dimensional volume to be inferred from multiple two-dimensional projections, is an important problem in imaging with applications in various domains including medical scanners, electron microscopy, non-destructive testing and baggage screening for security. Model-Based Iterative Reconstruction (MBIR) is a popular approach to 3D reconstruction that has demonstrated state-of-the-art reconstruction quality on several applications, and has been deployed in commercial healthcare systems as for its low dose requirement. However, software implementations of MBIR on commodity general-purpose processors demonstrate poor performance due to its high computation, data requirements and cache un-friendly data access patterns. Meanwhile, Computed Aided Detection has been developed with the rapid growth of computing capability and patterns recognition.The main research work consists of four following parts:(1) We analyze the 3D reconstruction system and use approximate computing to break the dependency from voxel to voxel between slices. We came up with the flow to ensure the quality of services and show the speedup of the approximate computing without the loss of quality in the reconstructed images.(2) We derive the model based iterative reconstruction algorithm efficiency by constraining the sequence in which voxels in the 3D volume are reconstructed. Employing this method, we develop an Efficient MBIR Accelerator (EMBIRA) that achieves significant performance and energy improvement over software implementations. EMBIRA utilizes arrays of 3 types of specialized processing elements that match MBIR’s computation patterns, and is further operated as a two-level nested pipeline to fully exploit the parallelism present in the algorithm. This enables better data reuse within the accelerator, thereby significantly reducing the number of off-chip memory accesses. To demonstrate the benefits of EMBIRA, we implemented a prototype on FPGA platform interfaced with an external DDR3 memory to show the detailed implementation optimization method using approximate computing and massive speedups and energy saving.(3) We employ the Haar feature cascade architecture to detect objects in the CT system after they are reconstructed to help the doctors for diagnosing. We develop a hardware system to detect different objects using different Haar feature library.(4) After analyzing the flow of the edge-orientation interpolation algorithm, we provide a topology based pipeline generation mechanism. And we verify the mechanism by improving and realize the cost effective interpolation algorithm based on edge-orientation map and realize the algorithm in hardware. The pipeline generation method can reduce great hardware cost by adjusting the interpolation flow and simplifying the topology of edge-orientation map.