PyHST2: an hybrid distributed code for high speed tomographic reconstruction with iterative reconstruction and a priori knowledge capabilities (1306.1392v1)

Abstract: We present the PyHST2 code which is in service at ESRF for phase-contrast and absorption tomography. This code has been engineered to sustain the high data flow typical of the third generation synchrotron facilities (10 terabytes per experiment) by adopting a distributed and pipelined architecture. The code implements, beside a default filtered backprojection reconstruction, iterative reconstruction techniques with a-priori knowledge. These latter are used to improve the reconstruction quality or in order to reduce the required data volume and reach a given quality goal. The implemented a-priori knowledge techniques are based on the total variation penalisation and a new recently found convex functional which is based on overlapping patches. We give details of the different methods and their implementations while the code is distributed under free license. We provide methods for estimating, in the absence of ground-truth data, the optimal parameters values for a-priori techniques.

• The paper introduces PyHST2, a hybrid distributed code designed for high-speed tomographic reconstruction at synchrotron facilities, capable of handling petabyte-scale data.
• PyHST2 integrates iterative reconstruction methods incorporating a-priori knowledge like Total Variation penalization and dictionary learning to improve image quality or reduce data requirements.
• The implementation leverages hybrid CPU/GPU architectures and parallel processing techniques for efficient data handling and high throughput, demonstrated through experimental applications.

Tomographic Reconstruction with PyHST2: An Examination of Hybrid Distributed Computing and A-priori Knowledge Approaches

The paper details the capabilities of PyHST2, a hybrid distributed code designed for high-speed tomographic reconstruction utilized at the European Synchrotron Radiation Facility (ESRF). This novel computational tool addresses the challenges of handling the vast data sets characteristic of modern synchrotron facilities, which can exceed 10 terabytes per experiment. PyHST2 leverages a distributed, pipelined architecture to enable efficient processing of these large volumes of data, particularly in the context of phase-contrast and absorption tomography.

In addition to implementing traditional filtered back-projection reconstruction, PyHST2 offers iterative reconstruction techniques that incorporate a-priori knowledge to enhance reconstruction quality or minimize data requirements. This paper emphasizes two significant a-priori techniques: total variation (TV) penalization and dictionary learning through overlapping patches. These approaches are crucial for improving image quality under conditions of noise or reduced data availability.

A-priori Knowledge Techniques

1. Total Variation Penalization: PyHST2 integrates total variation regularization, which is beneficial for reconstructing piecewise-constant images. The method promotes sparsity in the image gradient, which reduces noise artifacts and enhances image clarity. The paper describes the use of the iterative shrinkage-thresholding algorithm (ISTA) and its accelerated version (FISTA) to solve the resulting optimization problem. The choice of the regularization parameter, $\beta$, is crucial and typically requires calibration based on the noise level and the microstructural context of the data.
2. Dictionary Learning with Overlapping Patches: For images that are not piecewise constant, the total variation approach may not be optimal. Instead, PyHST2 employs a dictionary learning technique to construct an over-complete basis, allowing a sparse representation of the image patches. This method uniquely addresses the challenge of image denoising and reconstruction from limited data sets. The approach uses convex optimization techniques to ensure robustness and efficiency, with GPU acceleration facilitating computational performance.

Implementation and Performance

The implementation of PyHST2 involves a hybrid setup using multi-core processors and optional GPU accelerator cards. The system exploits parallel processing capabilities and memory management techniques to ensure efficient data handling and processing across distributed resources. The paper highlights the software's adaptability to existing computational infrastructure and the strategic use of hardware resources to maximize data throughput and minimize computational latency.

Experimental Applications

The efficacy of the presented techniques is demonstrated through several applications, including the reconstruction of quartz grain samples and synthetic phantoms. The paper provides insights into the parameter selection process, showcasing the use of statistical estimators, such as cross-validation and the novel decoherence maximizing estimator, to determine optimal parameter settings for a-priori reconstruction techniques.

Implications and Future Directions

PyHST2 exemplifies the convergence of high-performance computing and advanced reconstruction techniques, pushing the boundaries of synchrotron-based tomography. This work not only addresses practical challenges associated with large-scale data processing but also introduces theoretical advancements in tomographic reconstruction methods through the incorporation of a-priori knowledge.

The implications of this research extend to various domains where tomographic imaging is central, including medical imaging, materials science, and biological research. As detection technology evolves, PyHST2's adaptable and efficient architecture positions it to handle future demands, potentially incorporating additional a-priori methodologies and broader hardware support.

In conclusion, while the paper establishes PyHST2 as a significant tool in high-speed tomographic reconstruction, continuous improvements in a-priori algorithms and computational strategies will be necessary to keep pace with rapidly advancing technology and scientific inquiry.