University of Utah oneAPI Center of Excellence works on Developing Precise Real-Time Tumor Scanning
In August, the federal Advanced Research Projects Agency for Health (ARPA-H) announced up to $23 million in awards to develop new technologies for precise tumor removal under the current Biden-Harris administration's Cancer Moonshot program.
Tulane University and the University of Utah are leveraging training and machine learning, combined with the advanced, tightly error-controlled floating point compression of ZFP, to address a key challenge of cancer surgery:
Doctors want to ensure that as much of an affected organ as possible is preserved while at the same time also being confident that all cancerous cells have indeed been removed. The exact “margins” or boundaries of a cancerous mass are, however, very difficult to discern using non-invasive pre-op methods or even using direct microscopic observation by very experienced cancer treatment specialists. The consequences of even minor errors in judgment during surgery can be quite serious, and determining whether cancer removal was successful may take a long time, as the only definitive answer can be found if no new tumors are formed after an extended period of time.
Prof. Valerio Pascucci and his team at the Center for Extreme Data Management Analysis and Visualization (CEDMAV) and the Scientific Computing and Imaging Institute, and Price Engineering's Kahlert School of Computing are helping with the Machine-learning Assisted Gigantic-Image Cancer Margin Scanner (MAGIC-SCAN) project, which aims to address exactly this problem.
As part of the Cancer Moonshot program, they will create an imaging system that allows doctors to scan a tumor during surgery and determine in real-time whether any cancer tissue has been left behind. Adjustments can be made immediately, minimizing the need for repeated surgery and reducing patient risks. The surgeon will be able to ascertain and communicate whether removal was successful.
Check out the announcement by the University of Utah on the collaboration with Tulane University to overcome the technical computing and engineering challenges to make the advanced imaging device a reality within the next five years: University of Utah Collaborates with Tulane University on up to $23 Million Cancer Moonshot Project to Build Advanced Tumor Imaging System |
SYCL-Enabled ZFP Compression for Solution Scaling
Machine learning inference during surgery needs to be scalable to ensure that this new solution is not only available to cancer research hospitals with access to high-performance computing environments. The imaging data used for tissue analysis is in the order of petabytes during model training and terabytes for reliable inference. In short, this requires rapid processing of enormous amounts of data. This will consume a lot of the computational and memory bandwidth of the diagnostics system in the operating room.
Scalability from major medical research university hospitals to small community hospitals nationwide relies on state-of-the-art high-fidelity data preparation and data compression.
This is where the ZFP compression library, originally developer at Lawrence Livermore National Laboratory, comes to the rescue. It is an open-source library for compressed floating point and integer arrays. Works best for 2D, 3D, and 4D arrays that exhibit spatial correlation, such as continuous fields from physics simulations, natural images, or regularly sampled terrain surfaces. It enables tight data fidelity, compression loss controls, and efficient in-memory representation of multidimensional numerical data for computations, data analysis, and visualization.
As part of their contribution to the oneAPI Academic Center of Excellence program, Prof Valerio Pascucci and the research team have enabled ZFP to use the SYCL* cross-platform abstraction layer and components of the oneAPI software platform specification. This allows ZFP and MAGIC-SCAN to use an open standards based multivendor approach for solution scaling.
Using SYCL and having the full open software developer ecosystem of the Unified Acceleration Foundation (UXL) to back their efforts, it will be possible to achieve these scalability goals. It will be achievable to deploy MAGIC-SCAN on a large set of diverse hardware platform configurations, taking full advantage of parallel compute and accelerator offload on AI PC setups and mobile surgical equipment.
“The University of Utah, in collaboration with Tulane University, applies new funding of up to $23M by the U.S. President's Cancer Moonshot project and the Advanced Research Project Agency for Health (ARPA-H) to solve one of the central problems of tumor treatment and surgery: ensuring all diseased cells, especially at the “margins” of a cancerous mass are being removed. The goal of project MAGIC-SCAN is to revolutionize advanced AI-assisted imaging data analytics to determine within minutes whether any cancer tissue has been left behind, allowing doctors to make life-saving decisions reliably and in real-time. This requires fast processing of enormous amounts of data, in the order of petabytes during model training and terabytes for reliable inference in the operating room. State-of-the-art high-fidelity data preparation and data compression are key since the ultimate solution must also be deployed in rural hospitals without access to a large parallel computing environment. Open standards-based C++ with SYCL offload to diverse sets of GPUs and accelerators provides a high-performance backend to ZFP, a leading software for error-controlled compression. This ensures advanced data compression and throughput while maintaining the tight data quality control required. The Intel® oneAPI DPC++/C++ Compiler, together with the Intel® DPC++ Compatibility Tool and the Intel® VTune™ Profiler, the broad software ecosystem of oneAPI and the Unified Acceleration Foundation (UXL), allow to streamline development and integration. This will enable solution scaling across computing platforms, from specialized cancer treatment facilities in major medical centers to the operating rooms in small hospitals.” — Valerio Pascucci, Director, Center for Extreme Data Management Analysis and Visualization (CEDMAV), Professor, Scientific Computing and Imaging Institute, and Price Engineering's Kahlert School of Computing, University of Utah |
Using oneAPI and SYCL brings the benefits of ZFP compression and fast random data access to a potentially wide range of Intel and third-party hardware solutions, ensuring performance and scalability.
To port the codebase, the University of Utah used the Intel® DPC++ Compatibility Tool along with the GPU-aware Intel® VTune™ Profiler to migrate software from the proprietary vendor-locked CUDA* baseline with ease and optimize memory management and compute performance.
Check out the ZFP Compression Source Code and SYCL Support for it. |
Stay tuned for future updates on the progress of advanced machine learning assisted imaging for precise tumor removal as part of the Cancer Moonshot initiative. We look forward to scalable cancer treatments with predictably reliable positive health outcomes.
Your Project can Leverage SYCL and oneAPI Too
Join us on the journey towards open, scalable parallel compute, ML and AI with oneAPI and SYCL.
Get started today with the Intel® DPC++ Compatibility Tool, Intel® oneAPI DPC++ Compiler, and Intel® VTune™ Profiler.
Download them standalone or as part of the Intel® oneAPI Base Toolkit.
Additional Resources
Press Releases
- White House Remarks on ARPA-H Announcement on Technologies for Precise Tumor Removal [22:26]
- ARPA-H announces awards to develop novel technologies for precise tumor removal
- President Joe Biden will announce up to $23 million in funding for Tulane University to invent advanced cancer imaging system
- The U Collaborates with Tulane on up to $23 Million Cancer Moonshot Project to Build Advanced Tumor Imaging System
MAGIC-SCAN and ZFP
- Machine-learning Assisted Gigantic-Image Cancer Margin Scanner
- zfp: Compressed Floating-Point and Integer Arrays
- ZFP Compression Source Code
SYCL Migration
Rob enables developers to streamline programming efforts across multiarchitecture compute devices for high performance applications taking advantage of Intel's family of development tools. He has extensive 20+ years of experience in technical consulting, software architecture and platform engineering working in IoT, edge, embedded software and hardware developer enabling.