Developments in acquisition technology and a growing need for time-resolved experiments pose great computational challenges in tomography. In addition, access to reconstructions in real time is a highly demanded feature but has so far been out of reach. We show that by exploiting the mathematical properties of filtered backprojection-type methods, having access to real-time reconstructions of arbitrarily oriented slices becomes feasible. Furthermore, we present RECAST3D, software for visualization and on-demand reconstruction of slices. A user of RECAST3D can interactively shift and rotate slices in a GUI, while the software updates the slice in real time. For certain use cases, the possibility to study arbitrarily oriented slices in real time directly from the measured data provides sufficient visual and quantitative insight. Two such applications are discussed in this article.

Tomographic X-ray microscopy beamlines at synchrotron light sources worldwide have pushed the achievable time-resolution for dynamic 3-dimensional structural investigations down to a fraction of a second, allowing the study of quickly evolving systems. The large data rates involved impose heavy demands on computational resources, making it difficult to readily process and interrogate the resulting volumes. The data acquisition is thus performed essentially blindly. Such a sequential process makes it hard to notice problems with the measurement protocol or sample conditions, potentially rendering the acquired data unusable, and it keeps the user from optimizing the experimental parameters of the imaging task at hand. We present an efficient approach to address this issue based on the real-time reconstruction, visualisation and on-the-fly analysis of a small number of arbitrarily oriented slices. This solution, requiring only a single additional computing workstation, has been implemented at the TOMCAT beamline of the Swiss Light Source. The system is able to process multiple sets of slices per second, thus pushing the reconstruction throughput on the same level as the data acquisition. This enables the monitoring of dynamic processes as they occur and represents the next crucial step towards adaptive feedback control of time-resolved in situ tomographic experiments.

strongly connected to their 3D structure. [1][2][3][4][5][6][7][8][9][10] Heterostructures, in which several compounds are combined within a single nano-object, provide even more flexibility to tune their final properties. For example, bimetallic nanoparticles can display superior properties compared to their monometallic counterparts. [11][12][13] To understand the connection between structure/ composition and properties, nanoparticles are often investigated by transmission electron microscopy (TEM). Although TEM has become an indispensable tool for studying nanomaterials, it remains difficult to perform a 3D characterization. Indeed, conventional TEM provides 2D projection images of 3D objects, therefore missing a wealth of information. Electron tomography was developed to overcome this issue. [14][15][16][17] In 2003, Midgley et al. combined high angle annular dark field scanning TEM (HAADF-STEM) with tomography, [18,19] which has since been successfully applied to investigate a broad variety of nanostructures. [20][21][22][23][24] During a typical electron tomography experiment, a series of 2D projection images are collected along various tilt angles, to cover an angular range that is as large as possible. After alignment of the tilt series, they serve as the input to a mathematical algorithm that reconstructs the 3D structure of the object. Although the acquisition of a tilt series can be automated, it can take (many) hours to obtain all images, depending on the complexity of the experiment. In addition, both the alignment and the reconstruction of the acquired projection images are carried out through offline post-processing procedures, performed at a dedicated workstation. These steps are computationally demanding, leading to a total data processing time of at least 1 h. To dramatically accelerate the acquisition of tilt series, socalled "fast tomography" was recently introduced in both TEM and HAADF-STEM modes. [25][26][27] The methodology is based on continuously tilting the holder and simultaneously acquiring projection images, ideally while focusing and tracking the particle at the same time.Fast HAADF-STEM tomography enables a new range of experiments, during which the dynamic behavior of nanoparticles can be probed in 3D. For example, recently this technique was combined with in situ heating to investigate the thermal stability of Au and Au/Pd nanoparticles. [27,28] These experiments are at the state of the art with respect to acquisition time, and we were able to record a full HAADF-STEM tilt series within about 5 min. However, since the alignment A detailed 3D investigation of nanoparticles at a local scale is of great importance to connect their structure and composition to their properties. Electron tomography has therefore become an important tool for the 3D characterization of nanomaterials. 3D investigations typically comprise multiple steps, including acquisition, reconstruction, and analysis/quantification. Usually, the latter two steps are performed offline, at a dedicated workstation. This...

At x-ray beamlines of synchrotron light sources, the achievable time-resolution for 3D tomographic imaging of the interior of an object has been reduced to a fraction of a second, enabling rapidly changing structures to be examined. The associated data acquisition rates require sizable computational resources for reconstruction. Therefore, full 3D reconstruction of the object is usually performed after the scan has completed. Quasi-3D reconstruction—where several interactive 2D slices are computed instead of a 3D volume—has been shown to be significantly more efficient, and can enable the real-time reconstruction and visualization of the interior. However, quasi-3D reconstruction relies on filtered backprojection type algorithms, which are typically sensitive to measurement noise. To overcome this issue, we propose Noise2Filter, a learned filter method that can be trained using only the measured data, and does not require any additional training data. This method combines quasi-3D reconstruction, learned filters, and self-supervised learning to derive a tomographic reconstruction method that can be trained in under a minute and evaluated in real-time. We show limited loss of accuracy compared to training with additional training data, and improved accuracy compared to standard filter-based methods.

a b s t r a c tTomography is a powerful technique for 3D imaging of the interior of an object. With the growing sizes of typical tomographic data sets, the computational requirements for algorithms in tomography are rapidly increasing. Parallel and distributed-memory methods for tomographic reconstruction are therefore becoming increasingly common. An underexposed aspect is the effect of the data distribution on the performance of distributedmemory reconstruction algorithms. In this work, we introduce a geometric partitioning method, which takes into account the acquisition geometry and aims to minimize the necessary communication between nodes for distributed-memory forward projection and back projection operations. These operations are crucial subroutines for an important class of reconstruction methods. We show that the choice of data distribution has a significant impact on the runtime of these methods. With our novel partitioning method we reduce the communication volume drastically compared to straightforward distributions, by up to 90% for a number of cases, and furthermore we guarantee a specified load balance.

In tomography, reconstruction and analysis is often performed once the acquisition has been completed due to the computational cost of the 3D imaging algorithms. In contrast, real-time reconstruction and analysis can avoid costly repetition of experiments and enable optimization of experimental parameters. Recently, it was shown that by reconstructing a subset of arbitrarily oriented slices, real-time quasi-3D reconstruction can be attained. Here, we extend this approach by including realtime segmentation, thereby enabling real-time analysis during the experiment.We propose to use a convolutional neural network (CNN) to perform real-time image segmentation and introduce an adapted training strategy in order to apply CNNs to arbitrarily oriented slices. We evaluate our method on both simulated and real-world data. The experiments show that our approach enables realtime tomographic segmentation for real-world applications and outperforms standard unsupervised segmentation methods.

The bulk-synchronous parallel (BSP) programming model gives a powerful method for implementing and describing parallel programs. In this article we present Bulk, a novel interface for writing BSP programs in the C++ programming language that leverages modern C++ features to allow for the implementation of safe and generic parallel algorithms for shared-memory, distributed-memory, and hybrid systems. This interface targets the next generation of BSP programmers who want to write fast, safe, clear and portable parallel programs. We discuss two applications: regular sample sort and the fast Fourier transform, both in terms of performance, and ease of parallel implementation. IntroductionThe bulk-synchronous parallel (BSP) model was introduced as a bridging model for parallel programming by Valiant in 1989 [1]. It enables a way to structure parallel computations, which aids in the design and analysis of parallel programs.The BSP model defines an abstract computer, the BSP computer, on which BSP algorithms can run. Such a computer consists of p identical processors, each having access to their own local memory. A communication network is available which can be used by the different processors to communicate data. During the execution of an algorithm, there are points at which bulk synchronizations are performed. The time of such a synchronization, the latency, is denoted by l. The communication cost per data word is denoted by g. The parameters l and g are usually expressed in number of floating-point operations (FLOPs). They can be related to wall-clock time by considering the computation rate r of the individual processors which is measured in floating-point operations per second (FLOP/s). A BSP computer is captured completely by the parameter tuple (p, g, l, r).At a high level, a BSP algorithm is a series of supersteps that each consist of a computation phase and a communication phase. The processors of a BSP computer can simultaneously send and receive data, and they can do so independently. This means that the cost of communication is dominated by the maximum number of words sent or received by any processor. At the end of each superstep a bulk synchronization is performed ensuring that all outstanding

No abstract