HPCI 2025 Program
Excel to HTML
MONDAY 3 FEBRUARY 2025
Vision Foundation Models
Session Chair: Xiao Wang, Oak Ridge National Laboratory
09:50 - 10:30
Regency A
09:50HPCI-172
Write sentence with images: Revisit the large vision model with visual sentence, Quan Liu, Vanderbilt University; Can Cui, Vanderbilt University; Ruining Deng, Vanderbilt University; Tianyuan Yao, Vanderbilt University; Yucheng Tang, NVIDIA; Yuankai Huo, Vanderbilt University [view abstract]
This paper introduces a novel approach to image generation from visual sentences extracted from videos. By combining a lightweight autoregressive model with a Vector Quantized Generative Adversarial Network (VQGAN), we aim to bridge the gap between quality and computational efficiency. Unlike traditional methods, which often require extensive computational resources, our approach achieves comparable performance to state-of-the-art models while improving processing efficiency. The autoregressive model captures the sequential patterns within the visual sentences, allowing for more coherent and contextually accurate image generation. Our experimental results demonstrate that this approach can generate high-quality images with a lower computational burden, making it a viable option for applications requiring real-time or resource-constrained environments. This work presents a balance between performance and efficiency that could be beneficial in various multimedia and creative domains.
10:10HPCI-173
Challenges and methods in kilometer-scale E3SM land model simulation data processing and Analysis, Dali Wang, Oak Ridge National Laboratory, (US); Peter Schwartz, Oak Ridge National Laboratory, (US); Fengming Yuan, Oak Ridge National Laboratory, (US); Peter Thornton, Oak Ridge National Laboratory, (US); Qinglei Cao, Saint Louis University (US); Chen Wang, Lawrence Livermore National Laboratory, (US); Xiao Wang, Oak Ridge National Laboratory, (US) [view abstract]
The increasing need for detailed insights into Earth's climate dynamics has prompted the use of kilometer-scale Earth System Models (ESMs), which provide high-resolution simulations critical for understanding both natural and anthropogenic influences on the climate system. However, such fine-scale simulations pose significant challenges, primarily due to their substantial computational demands and the complexity in managing and analyzing vast amounts of output data. This presentation focuses on the technical hurdles encountered while processing ultrahigh-resolution simulation data, specifically results from 1km x 1km resolution simulations over North America. We explore innovative approaches, including advanced image processing techniques and machine learning methodologies, aimed at efficiently handling and interpreting the intricate data generated. By leveraging these modern analytical tools, we seek to enhance our capacity to extract meaningful insights from ESM outputs. Ultimately, this discussion aims to bridge the gap between experts in image processing and those engaged in Earth system simulations, fostering collaboration that may lead to improved data analysis strategies and a deeper understanding of climate change phenomena.
10:10HPCI-345
ExtremeMETA: High-speed lightweight image segmentation model by remodeling multi-channel metamaterial imagers, Quan Liu, Vanderbilt University and Oak Ridge National Laboratory (US); Brandon Swartz, Vanderbilt University and Oak Ridge National Laboratory (US); Ivan Kravchenko, Vanderbilt University and Oak Ridge National Laboratory (US); Jason Valentine, Vanderbilt University and Oak Ridge National Laboratory (US); Yuankai Huo, Vanderbilt University and Oak Ridge National Laboratory (US) [view abstract]
Deep neural networks (DNNs) have heavily relied on traditional computational units, such as CPUs and GPUs. However, this conventional approach brings significant computational burden, latency issues, and high power consumption, limiting their effectiveness. This has sparked the need for lightweight networks such as ExtremeC3Net. Meanwhile, there have been notable advancements in optical computational units, particularly with metamaterials, offering the exciting prospect of energy-efficient neural networks operating at the speed of light. Yet, the digital design of metamaterial neural networks (MNNs) faces precision, noise, and bandwidth challenges, limiting their application to intuitive tasks and low-resolution images. In this study, we proposed a large kernel lightweight segmentation model, ExtremeMETA. Based on ExtremeC3Net, our proposed model, ExtremeMETA maximized the ability of the first convolution layer by exploring a larger convolution kernel and multiple processing paths. With the large kernel convolution model, we extended the optic neural network application boundary to the segmentation task. To further lighten the computation burden of the digital processing part, a set of model compression methods was applied to improve model efficiency in the inference stage. The experimental results on three publicly available datasets demonstrated that the optimized efficient design improved segmentation performance from 92.45 to 95.97 on mIoU while reducing computational FLOPs from 461.07 MMacs to 166.03 MMacs. The large kernel lightweight model ExtremeMETA showcased the hybrid design's ability on complex tasks.
10:30 – 11:00 and 15:00 - 15:30 Coffee Breaks
Efficient Tools & Infrastructure for Imaging
Session Chair: Yuankai Huo, Vanderbilt University
15:30 - 17:30
Regency A
15:30HPCI-174
Infrastructure management through the fusion of artificial intelligence and high performance imaging, Du Wu, TokyoTech, (Japan); Peng Chen, RIKEN Center for Computational Science, (Japan); Enzhi Zhang, Hokkaido University, (Japan); Jun Igarashi, RIKEN Center for Computational Science, (Japan); Issac Lyngaas, ORNL, (US); Xiao Wang, ORNL, (US); Mohamed Wahib, RIKEN Center for Computational Science, (Japan) [view abstract]
This presentation explores the integration of Artificial Intelligence (AI) with High Performance Imaging (HPI) to enhance road infrastructure management in Japan. By leveraging advanced AI techniques such as machine learning and computer vision, alongside cutting-edge imaging technologies, we address critical challenges in monitoring, analyzing, and maintaining large-scale infrastructures. The fusion of these technologies enables detection of structural issues, predictive maintenance, and resource optimization. We will highlight the transformative potential of AI and HPI in improving safety, efficiency, and sustainability across sectors such as construction, energy, and transportation. The discussion will highlight innovations in data processing, image analysis, and decision-making frameworks, emphasizing how this interdisciplinary approach can revolutionize infrastructure management.
15:50HPCI-175
SparVNM: Efficient vector-wise N:M sparsity implementation for GPGPU, Cong Ma, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, (China (Mainland)); Xiaowen Huang, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, (China (Mainland)); Xu Zhang, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, (China (Mainland)); Jintao Meng, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, (China (Mainland)) [view abstract]
In image processing, where DNNs are widely used for tasks such as image classification, object detection, and semantic segmentation, large models face significant computational overhead, limiting their deployment on resource-constrained devices. Weight pruning in Deep Neural Networks (DNNs) has been extensively studied to reduce model size and accelerate inference. Pruning techniques are classified into structured, unstructured, and the more recent semi-structured N:M sparsity. Consequently, reducing model size and computational requirements without sacrificing accuracy is crucial. Unstructured pruning introduces irregular computation and memory access patterns, making practical GPU acceleration challenging, while structured pruning often struggles to balance performance and speedup on these platforms. N:M sparsity has emerged as a promising alternative, enforcing a constraint to retain N out of every M units, though Sparse Tensor Core support is currently limited to 2:4 sparsity. To address this, nmSPARSE, a GPU library of SpMM kernels, was developed to support general N:M sparsity. We further propose SparVNM, an efficient vector-wise N:M sparsity implementation using multi-layer tiling and double buffering to optimize memory access. Experiments show SparVNM achieves a 2.1x speedup over nmSPARSE and a 1.4x to 6.3x speedup over dense cuBLAS, approaching ideal speedup from sparsity.
16:10HPCI-176
Synthetic dataset pre-training for precision medical segmentation using vision transformers, Edgar Josafat Martinez-Noriega, AIST (Japan); Rio Yokota, Tokyo Institute of Technology Blog About (Japan); Peng Chen, AIST, (Japan); Thao Nguyen Truong , AIST (Japan) [view abstract]
In medical segmentation, the acquisition of high-quality labeled data remains a significant challenge due to the substantial cost and time required for expert annotations. Variability in imaging conditions, patient diversity, and the use of different imaging devices further complicate model training. The high dimensionality of medical images also imposes considerable computational demands, while small lesions or abnormalities can create class imbalance, thus hindering segmentation accuracy. Pre-training on synthetic datasets in medical imaging may enable Vision Transformers (ViTs) to develop robust feature representations, even in the fine-tuning phase, when high-quality labeled data is limited. In this work, we propose integrating Formula-Driven Supervised Learning (FDSL) synthetic datasets with medical imaging to enhance pre-training for segmentation tasks. We implemented a custom Fractal dataset capable of generating high-resolution images, including those measuring 64K x 64K pixels or larger. Preliminary results indicate improved performance when using the SAM model for segmentation in conjunction with robust augmentation techniques, followed by fine-tuning on the PAIP dataset, a high-resolution, real-world pathology dataset focused on liver cancer. Additionally, we present results using another synthetic dataset, SegRCDB, for comparative analysis.
16:30HPCI-177
Pyspatial: A high-speed whole slide image pathomics toolkit, Yuechen Yang, Department of Computer Science, Vanderbilt University, Nashville, TN, US; Tianyuan Yao, Department of Computer Science, Vanderbilt University, Nashville, TN, US; Ruining Deng, Department of Computer Science, Vanderbilt University, Nashville, TN, US; Haichun Yang, Department of Pathology, Microbiology and Immunology, Vanderbilt University Medical Center, Nashville, TN, US; Yuankai Huo, Department of Computer Science, Vanderbilt University, Nashville, TN, US [view abstract]
In this paper, we present PySpatial, a new open-source pathomics toolkit optimized for Whole Slide Image (WSI) analysis. PySpatial addresses the challenges of processing super high-resolution histopathological data by efficiently differentiating giga-pixel images into computational and non-computational regions. This approach not only enhances computational speed but also minimizes memory USge by focusing primarily on the essential computational areas. A novel data structure, the Computational Region Pool (CRP), has been developed to efficiently store and process only relevant image segments. By offering PySpatial as an open-source resource, we aim to empower researchers to perform WSI-level pathomic analyses more effectively and accurately.
16:50HPCI-178
The libvips image processing library, Kirk Martinez, University of Southampton, (United Kingdom); John Cupitt; Lovell Fuller; Kleis Wolthuizen [view abstract]
libvips is a LGPL licensed (open source and free for commercial use), portable, horizontally-threaded, demand-driven, 2D image processing library with its origins in imaging research projects. Compared to similar libraries, libvips runs quickly and uses little memory. It supports numeric formats from 8-bit integer to 128-bit complex, any number of color separation bands, most popular image formats, and many specialized scientific image formats. It has become popular in applications such as virtual microscopy and art imaging, and very popular as an image processing library for the web. This paper outlines the history of the library, explains how libvips achieves its good performance, presents benchmarks, and gives an overview of the implementation and of the wider libvips ecosystem.
17:10HPCI-179
Steady-state particle advection speedups from GPU and CPU parallelism, Abhishek Yenpure, Kitware Inc., (US); Hank Childs, University of Oregon, (US); David Pugmire, Oak Ridge National Laboratory, (US) [view abstract]
This study evaluates the benefit of using parallelism from GPUs or multi-core CPUs for particle advection workloads. We perform 1000+ experiments, involving four generations of Nvidia GPUs, four CPUs with varying numbers of cores, two particle advection algorithms, many different workloads (i.e., number of particles and number of steps), and, for GPU tests, performance with and without data transfer. The results inform whether or not a visualization developer should incorporate parallelism in their code, what type (CPU or GPU), and the key factors influencing performance. Finally, we find that CPU parallelism is the better choice for most common workloads, even when ignoring costs for data transfer.
TUESDAY 4 FEBRUARY 2025
KEYNOTE Session
Session Chair: Yuankai Huo, Vanderbilt University
09:00 - 10:30
Regency A
09:00HPCI-180
Boosting ultrasound computed tomography with generative AI and neural PDE solvers, Zhijun , Tsinghua University; Youjia, Peking University; Zeyuan ; Yubing He, Peking University (China) [view abstract]
Ultrasound Computed Tomography (USCT) offers a radiation-free option for high-resolution clinical imaging, but its widespread adoption is limited by computational demands and numerical instability in its image reconstruction. This process requires solving a Partial Differential Equation (PDE)-constrained inverse problem known as Full Waveform Inversion (FWI), which typically involves costly multiple iterations of numerical wave equation solvers. In this paper, we introduce a Real2Sim2Real approach that enhances wave simulations and FWI-based reconstruction for USCT through the use of generative AI and neural PDE solvers. Our method begins with generative AI to construct a synthetic dataset of over ten million pairs of anatomically realistic phantoms and their corresponding multi-frequency wavefields from various organs (Real2Sim). We then developed the Strong Scattering Neural Operator (S2NO), a novel neural surrogate model that accurately maps scattering medium parameters to wavefield solutions, and finally integrated it into clinical USCT FWI reconstruction (Sim2Real). We validated this framework through high-resolution wave simulations and imaging of both soft and bone tissues. Our experiments demonstrated its capability to achieve near real-time 3D imaging of human breasts, legs, and arms in under five minutes� producing medical images comparable to those of gold-standard modalities (e.g., MRI) while achieving an order-of-magnitude speedup over traditional FWI solvers.
09:40HPCI-181
Scaling vision transformers to 113 billion parameters with ORBIT and transforming climate modeling, Xiao Wang, Oak Ridge National Laboratory, (US) [view abstract]
Earth system predictability is challenged by the complexity of environmental dynamics and the multitude of variables involved. Current AI foundation models, although advanced by leveraging large and heterogeneous data, are often constrained by their size and data integration, limiting their effectiveness in addressing the full range of Earth system prediction challenges. To overcome these limitations, we introduce the Oak Ridge Base Foundation Model for Earth System Predictability (ORBIT), an advanced vision transformer model that scales up to 113 billion parameters using a novel hybrid tensor-data orthogonal parallelism technique. As the largest model of its kind, ORBIT surpasses the current climate AI foundation model size by a thoUSndfold. Performance scaling tests conducted on the Frontier supercomputer have demonstrated that ORBIT achieves 684 petaFLOPS to 1.6 exaFLOPS sustained throughput, with scaling efficiency maintained at 41% to 85% across 49,152 AMD GPUs. These breakthroughs establish new advances in AI-driven climate modeling and demonstrate promise to significantly improve the Earth system predictability.
10:30 – 11:00 Coffee Break
Special HPCI Session I: High-Efficient Computation for Medical Image Analysis
Session Chair: Yuankai Huo, Vanderbilt University
11:00 - 12:20
Regency A
11:00HPCI-182
Variational patches for high-resolution medical image segmentation with vision transformers, Enzhi Zhang, Hokkaido University, (Japan); Isaac Lyngaas, Oak Ridge National Laboratory, (US); Peng Chen, RIKEN Center for Computational Science, (Japan); Xiao Wang, Oak Ridge National Laboratory, (US); Yuankai Huo, Vanderbuilt University (US); Masaharu Munetomo, RIKEN Center for Computational Science, (Japan); Mohamed Wahib, RIKEN Center for Computational Science, (Japan) [view abstract]
Attention-based models have become increasingly popular in image analytics tasks such as segmentation. Typically, images are divided into patches and fed into transformer encoders as linear sequences of tokens. However, for high-resolution images, such as microscopic pathology images, the quadratic growth of computational and memory requirements makes the use of attention-based models challenging, especially when smaller patch sizes are needed for accurate segmentation. Existing solutions involve either complex multi-resolution models or approximate attention mechanisms. In this work, we propose a novel approach inspired by Adaptive Mesh Refinement (AMR) methods used in high-performance computing. Our method adaptively selects patches based on image details, significantly reducing the number of patches while maintaining fine-grained segmentation accuracy. This pre-processing technique introduces minimal overhead and can be seamlessly integrated with any attention-based model. We demonstrate improved segmentation performance on real-world pathology datasets, achieving a geometric mean speedup of $6.9\times$ for images up to $64K^2$ resolution, utilizing up to $2,048$ GPUs, compared to state-of-the-art segmentation models.
11:20HPCI-183
mTREE: Multi-level text-guided representation end-to-end learning for whole slide image analysis, Quan Liu; Ruining Deng; Can Cui; Tianyuan Yao; Vishwesh Nath ; Bingshan Li; You Chen; Yucheng Tang ; Yuankai Huo [view abstract]
Multi-modal learning adeptly integrates visual and textual data, but its application to histopathology image and text analysis remains challenging, particularly with large, high-resolution images like gigapixel Whole Slide Images (WSIs). Current methods typically rely on manual region labeling or multi-stage learning to assemble local representations (e.g., patch-level) into global features (e.g., slide-level). However, there is no effective way to integrate multi-scale image representations with text data in a seamless end-to-end process. In this study, we introduce Multi-Level Text-Guided Representation End-to-End Learning (mTREE). This novel text-guided approach effectively captures multi-scale WSI representations by utilizing information from accompanying textual pathology information. mTREE innovatively combines � the localization of key areas (global-to-local) and the development of a WSI-level image-text representation (local-to-global) � into a unified, end-to-end learning framework. In this model, textual information serves a dual purpose: firstly, functioning as an attention map to accurately identify key areas, and secondly, acting as a conduit for integrating textual features into the comprehensive representation of the image. Our study demonstrates the effectiveness of mTREE through quantitative analyses in two image-related tasks: classification and survival prediction, showcasing its remarkable superiority over baselines. Code and trained models are made available at https://github.com/hrlblab/mTREE.
11:40HPCI-184
Scale-up unlearnable examples learning with high-performance computing, Yanfan Zhu, (China (Mainland)); Issac Lyngaas; Murali Meena; Mary Ellen Koran; Bradley Malin; Daniel Moyer; Shunxing Bao; Anuj Kapadia; Bennett Landman; Xiao Wang; Yuankai Huo [view abstract]
Recent advancements in AI models, like ChatGPT, are structured to retain user interactions, which could inadvertently include sensitive healthcare data. In the healthcare field, particularly when radiologists use AI-driven diagnostic tools hosted on online platforms, there is a risk that medical imaging data may be repurposed for future AI training without explicit consent, spotlighting critical privacy and intellectual property concerns around healthcare data USge. Addressing these privacy challenges, a novel approach known as Unlearnable Examples (UEs) has been introduced, aiming to make data unlearnable to deep learning models. A prominent method within this area, called Unlearnable Clusters (UCs), has shown improved UE performance but was previously limited by computational resources (e.g., a single workstation) which restricted the exploration of larger-scale experiments.To push the boundaries of UE performance with theoretically vast resources, we scaled up UCs learning across various datasets using Distributed Data Parallel (DDP) training on the Summit supercomputer. Our goal was to examine UE efficacy at high-performance computing (HPC) levels to prevent unauthorized learning and enhance data security, particularly focusing on the influence of smaller batch sizes in increasing unlearnability, reflected by reduced model accuracy.Through the computational power of Summit, extensive testing on diverse datasets�such as Pets, MedMNist, Flowers, and Flowers102�was conducted. Results showed that excessively large or small batch sizes led to unstable performance and influenced accuracy. Smaller batch sizes on specific datasets such as pathMNist, BloodMNist and Flowers102 generally correlated with lower accuracy, a desired outcome for unlearnable data, shielding it from inference attacks. However, the optimal batch size for unlearnability varied across datasets, underscoring the need for dataset-specific computational scaling for effective data protection. Summit's high-performance GPUs, paired with DDP efficiency, enabled rapid parameter updates and consistent training across nodes, essential for determining batch sizes that maximize unlearnability without sacrificing computational efficiency. These findings highlight the importance of selecting suitable batch sizes tailored to dataset characteristics to prevent unauthorized model learning and ensure data security in deep learning applications. The source code for this study is publicly accessible at https://github.com/hrlblab/UE_HPC.
12:00HPCI-351
Segment anything model (SAM) for digital pathology: Assess zero-shot segmentation on whole slide imaging, Ruining Deng, Vanderbuilt University (US); Can Cui, Vanderbuilt University (US); Quan Liu, Vanderbilt University (US); Tianyuan Yao, Vanderbilt University; Lucas Remedios, Vanderbuilt University (US); Shunxing Bao, Vanderbuilt University (US); Bennett Landman, Vanderbilt University (US); Lee Wheless, Vanderbilt University Medical Center (US); Lori Coburn, Vanderbilt University Medical Center (US); Keith Wilson, Vanderbilt University Medical Center, (US); Hongyao Wang, Vanderbilt University Medical Center (US); Shilin Zhao, Vanderbilt University Medical Center (US); Agnes Fogo, Vanderbilt University Medical Center (US); Haichun Yang, Vanderbilt University Medical Center (US); Yucheng Yang, NVIDIA (US); Yuankai Huo, Vanderbilt University (US) [view abstract]
The segment anything model (SAM) was released as a foundation model for image segmentation. The promptable segmentation model was trained by over 1 billion masks on 11M licensed and privacy-respecting images. The model supports zero-shot image segmentation with various segmentation prompts (e.g., points, boxes, masks). It makes the SAM attractive for medical image analysis, especially for digital pathology where the training data are rare. In this study, we evaluate the zero-shot segmentation performance of SAM model on representative segmentation tasks on whole slide imaging (WSI), including (1) tumor segmentation, (2) non-tumor tissue segmentation, (3) cell nuclei segmentation. Core Results: The results suggest that the zero-shot SAM model achieves remarkable segmentation performance for large connected objects. However, it does not consistently achieve satisfying performance for dense instance object segmentation, even with 20 prompts (clicks/boxes) on each image.We also summarized the identified limitations for digital pathology: (1) image resolution, (2) multiple scales, (3) prompt selection, and (4) model fine-tuning. In the future, the few-shot fine-tuning with images from downstream pathological segmentation tasks might help the model to achieve better performance in dense object segmentation.
15:00 – 15:30 Coffee Break
Special HPCI Session II: Computing Infrastructure for Democratizing High-Performance Imaging
Session Chair: Yuhao Zhu, University of Rochester
15:30 - 17:30
Regency A
15:30HPCI-185
System and algorithm co-design for efficient image rendering in AR/VR system, [view abstract]
The cost of image rendering in AR/VR environments is considerable, primarily due to the demands of high-quality visual experiences from users. This challenge is even greater in real-time applications, where maintaining low latency further increases the complexity of the rendering process. On the other hand, VR devices, such as head-mounted displays (HMDs), are intrinsically linked to human behavior, using insights from perception and cognition to enhance user experience.In this talk, I will discuss recent research outcomes from my group that aim to reduce the high computational costs of the rendering process in VR. By leveraging the natural dynamics of human eye behavior and co-optimizing AI algorithms with the underlying supporting hardware architectures, we can achieve significant efficiency improvement on VR system performance.
16:10HPCI-186
Temporal architectures for vision sensor processing, Ryan Hou, University of Michigan, (US); Thomas Twomey; Vasileios Milionis; Georgios Tzimpragos, University of Wisconsin-Madison (US) [view abstract]
Binary representation and Boolean logic have long been the standard for general-purpose computing and storage, butit remains unclear if they are always the optimal choice. We posit that digital temporal codes, combined with a new class of temporal logic, present a unique advantage for near- and in-sensor processing.Temporal codes use time as a resource, encoding data values as delays in discrete voltage transitions, while logic is performed by engineering "race" conditions among these transitions.This approach addresses the scalability and noise issues of pure analog systems, as well as the costly analog-to-digital conversions and limited innovation space of traditional digital systems, all while maintaining full compatibility with off-the-shelf CMOS components for enhanced practicality.Building on our previous work, which demonstrated that temporal XGBoost accelerators achieved orders-of-magnitude improvements in energy-delayproductover state-of-the-art solutions, in this talk, wewill discuss how temporal computing can bring similar benefits to vision sensor interfaces, with extensions to VIO/SLAM pipelines.
16:50HPCI-187
State-space models for imaging at the edge, Mel White, Rice University (US) [view abstract]
SSMs have (re-)emerged in recent years as a powerful tool for online compression and efficient computing. However, the current published work is limited to a few orthonormal bases and continuous input signals, which has limited their implementation in imaging systems. Building on ongoing work to expand the SSM toolkit, we demonstrate how wavelet frames in particular can be used with SSMs for image compression and recognition tasks.
WEDNESDAY 5 FEBRUARY 2025
Software Landscape for Computational Imaging
Session Chair: Xiao Wang, Oak Ridge National Laboratory
09:10 - 10:30
Regency B
09:50HPCI-188
tomoORNL: Computed tomography algorithms for scientific imaging, Singanallur Venkatakrishnan, Oak Ridge National Laboratory, (US); Obaidullah Rahman; Chen Zhang ; Jean Bilheux; Amirkoushyar Ziabari [view abstract]
In this work, we will present tomoORNL, a computed tomography (CT) reconstruction library developed at Oak Ridge National Laboratory. tomoORNL has a broad selection of algorithms for CT simulation, pre-processing and reconstruction. A specific focus of the library has been to make Model-Based Iterative Reconstruction methods available in practice for scientific CT systems. tomoORNL has enabled prototyping of CT algorithms with experimental data corresponding to large volumes (1k X 1k X 1K and larger) using multi-GPU systems. We will demonstrate the value of the library by highlighting its use in applications including electron tomography, cryo-EM recosntruction, cone-beam X-ray CT, neutron tomography and laminography.
10:30 – 11:00 Coffee Break
Computing Efficient Tomography Imaging
Session Chair: Mohamed Wahib, RIKEN
11:00 - 12:20
Regency A
11:00HPCI-189
High-performance Image Reconstruction using ABCI supercomputer, Peng Chen, National Institute of Advanced Industrial Science and Technology AIST (Japan); Wahib Mohamed, RIKEN-CCS, (Japan); Xiao Wang, ORNL, (US); Isaac Lyngaas, ORNL, (US); Jun Igarashi, RIKEN-CCS, (Japan) [view abstract]
In this work, we showcase our high-performance image reconstruction framework utilizing the ABCI supercomputer. Computed Tomography (CT) is a critical imaging technique, but its image reconstruction algorithms are computationally intensive. To address this challenge, we introduce a novel back-projection algorithm that reduces projection computation costs compared to traditional methods. Our solution efficiently takes advantage of GPU-accelerated systems by dividing the workload between CPUs and GPUs, overlapping the filtering stage on CPUs with the back-projection on GPUs to optimize performance. Additionally, we present a distributed framework designed for high-resolution image reconstruction on modern supercomputers. This framework optimizes communication by interleaving MPI collective operations, enabling scalable data transfer. By scaling across up to 2,048 V100 GPUs on the ABCI supercomputer, we demonstrate near-instantaneous CT image reconstruction, showcasing the scalability and real-time capability of our approach.
11:20HPCI-190
Physics-guided AI/ML models for ptychographic image reconstruction, [view abstract]
While the advances in synchrotron radiation sources, together with the development of focusing optics and detectors, allow nanoscale ptychographic imaging of materials and biological specimens, the corresponding experiments can yield terabyte-scale volumes of data with heavy computational demands and long data acquisition times.In this talk, I will introduce how AI/ML models can be integrated into ptychographic reconstruction tasks to reduce probe overlap requirements, thereby significantly decreasing both data acquisition and computational costs. Specifically, I will present two methods: a scalable ptychographic reconstruction technique withuncertainty estimates, and an image reconstruction method that uses a score-based diffusion model guided by ptychographic imaging physics. Our experimental evaluations show that these methods consistently yield high-quality image reconstructions, even with extremely low probe overlaps.
11:40HPCI-191
High-performance tomographic imaging powered by tensor-core and PTX assembly, Du Wu, RIKEN-CCS, (Japan); Peng Chen, National Institute of Advanced Industrial Science and Technology, (Japan); Xiao Wang, Oak Ridge National Laboratory, (US); Issac Lyngaas, Oak Ridge National Laboratory; Takaaki Miyajima, Meiji University; Toshio Endo, Tokyo Institute of Technology; Satoshi Matsuoka, RIKEN-CCS; Mohamed Wahib, RIKEN-CCS, (Japan) [view abstract]
Computed Tomography (CT) serves as a key imaging technology that relies on computationally intensive filtering and back-projection algorithms for 3D image reconstruction. While conventional high-resolution image reconstruction (> 23) solutions provide quick results, they typically treat reconstruction as an offline workload to be performed remotely on large-scale HPC systems. The growing demand for post-construction AI-driven analytics and the need for real-time adjustments call for high-resolution reconstruction solutions that are feasible on local computing resources, i.e. on a single GPU server. In this paper, we propose a novel approach that utilizes Tensor Cores to optimize image reconstruction without sacrificing precision. Evaluations conducted on a single Nvidia A100 and H100 GPU show performance improvements of 1.91� and 2.15� compared to highly optimized production libraries.
12:00HPCI-192
SinoTx: A transformer-based model for sinogram inpainting, Jiaze E; Zhengchun Liu; Tekin Bicer; Srutarshi Banerjee; Rajkumar Kettimuthu; Bin Ren; Ian Foster [view abstract]
Sinogram inpainting is a critical task in computed tomography (CT) imaging, where missing or incomplete sinograms can significantly decrease image reconstruction quality. High-quality sinogram inpainting is essential for achieving high-quality CT images, enabling better diagnosis and treatment. To address this challenge, we propose SinoTx, a model based on the Transformer architecture specifically designed for sinogram completion. SinoTx leverages the inherent strengths of Transformers in capturing global dependencies, making it well-suited for handling the complex patterns present in sinograms. Our experimental results demonstrate that SinoTx outperforms existing baseline methods, achieving up to a 32.3% improvement in the Structural Similarity Index (SSIM) and a 44.2% increase in Peak Signal-to-Noise Ratio (PSNR).