Ultrasound has become one of the most common imaging modalities in the past two decades in fields such as sonar, biomedical imaging, and non-destructive testing (NDT). Ultrasonic waves in NDT are used to inspect and evaluate the integrity of different parts and materials, from rockets and jet engines to critical infrastructure like pipes and wind turbines. Computer vision and image analysis are routinely used to detect and characterize flaws, cracks, and other hazardous or dangerous defects. Early detection of issues and dangers in components helps save billions of dollars annually worldwide, and prevents catastrophic impacts to life and the environment. These methods help ensure the safe operation of aircrafts, concrete structures, railways, and critical infrastructure in energy production that we rely on as a society.
Computer vision techniques in NDT have advanced quickly in the past few years. While traditional techniques remain useful, modern ultrasound devices can easily collect vast quantities of high-resolution data in a short amount of time, making them approachable by deep learning methods. For example, a continuous acoustic scan at sub-millimetric resolutions of the full circumference of a pipe that is multiple kilometers long can yield millions of images in a few hours. Such datasets may contain their own unique challenges, including extreme data imbalance, or the need for multi-task, weakly-supervised and semi-supervised learning. In addition, gaps remain between natural image-derived deep learning algorithms, and those for ultrasonic acoustic-derived images, including focused image denoising, image interpretation, uncertainty quantification, and automated system self-awareness. Research into these topics is therefore important for developing robust deep-learning applications in NDT.
In the last few years, the medical ultrasound field has witnessed the successful application of deep learning in both in 2D and 3D to enhance, identify, and significantly speed up the analysis process. However, in the field of NDT, ultrasound analysis comes with its own set of challenges. For example, the imaging media and conditions can be significantly less predictable than the human body, and therefore many medical ultrasound analysis methods cannot be applied immediately to NDT ultrasound data. Therefore, methods specifically for the NDT domain are thus of paramount importance to keep up with the increasingly complex demands in NDT.
The Ultrasound Dataset Challenge
In this workshop, we invite prominent researchers at the interface of NDT and deep learning to explore the future of the field. We propose a challenge of 3D surface mesh estimation. Individual researchers and teams familiar with deep learning for visual tasks are welcome to register. The required context for this challenge is explained alongside its accompanying dataset. This unprecedented NDT ultrasound dataset, released for the first time, contains labels and annotations which allow it to be used for multi-task, semi-supervised, and weakly-supervised deep learning techniques, among others.
Workshop Date: June 18, 2023
08:20 am: Opening Remarks
"Overview of AI/ML as applicable to NDT"
09:00 am – 09:30 am: Paul D. Wilcox – Bristol University:
"Opportunities and Challenges for Deep Learning in Ultrasonic NDE Applications"
09:30 am – 10:00 am: Reza Zahiri and Jason Vantomme – DarkVision Technologies Inc.:
"From Sound Waves to Intelligent Solutions: Unveiling the Synergy of Ultrasound and Machine Learning"
10:00 am – 10:15 am: coffee break
Oral Paper Presentations
10:15 am – 10:30 am: Nick Luiken – KAUST:
"A deep learning-based approach to increase efficiency in the acquisition of ultrasonic non-destructive testing datasets"
10:30 am – 10:45 am: Blake VanBerlo – University of Waterloo:
"Exploring the Utility of Self-Supervised Pretraining Strategies for the Detection of Absent Lung Sliding in M-Mode Lung Ultrasound"
10:45 am – 11:00 am: Abder-Rahman Ali – Massachusetts General Hospital/Harvard Medical School:
"Self-Supervised Learning for Accurate Liver View Classification in Ultrasound Images with Minimal Labeled Data"
11:00 am – 11:15 am: Sourabh Kulhare – Global Health Labs, LLC:
"Deep Learning Video Classification of Lung Ultrasound Features Associated with Pneumonia"
11:15 am – 11:25 am: Challenge Introduction – Gareth Munro, DarkVision Technologies Inc.
11:25 am – 11:35 am: Lisa Tang – UBC:
"SMRVIS: Point cloud extraction from 3-D ultrasound for non-destructive testing"
11:35 am – 11:45 am: Yizhe Liu – University of Waterloo:
"CVPR 2023 Ultrasonic Data Challenge Project Report"
11:45 am: Closing Remarks
11:00 am – 12:00 pm: Poster session
In recent days a lot of emphasis is given on Machine learning techniques and now this has become prevalent in the area of Non-Destructive Testing (NDT)as well. Programs with Machine Learning (ML) capabilities are available in the market for automated discontinuity identification from data obtained from NDE sensors. It is important for such programs and functionalities to be qualified prior to use for these applications. It is necessary for the end users to understand the basics of Machine learning and how it works and the importance of data. ASNT has now formed a sub-committee for AI/ML that will take the lead to educate the Industry on AI/ML. It will use all available platforms for messaging to bring the brightest experts to the table to create tutorial publications and develop standards. This will also include engaging academic and research partners, development of new levels of personnel certifications and training programs.
This talk will first discuss the wider context of ultrasonic NDE and the resulting environment in which deep learning approaches must operate. The physical contrast between ultrasound in NDE and medical applications will be briefly noted. Several ultrasonic deep learning application examples from the author’s research group will then be considered, including measurement of a structural property, defect detection, and defect characterisation. Common themes that emerge from these include the need for precise problem definitions, accurate physics-based forward models, and objective assessments of performance. Possible ways of quantifying uncertainty and mitigating the shortage of labelled, true-positive training data will be suggested. Finally, future opportunities and research directions will be discussed.
This presentation explores the integration of ultrasound technology and machine learning for non-destructive testing in industrial applications. We examine the foundational role of ultrasound in providing non-invasive, real-time imaging for inspecting industrial materials and systems. By delving into ultrasound imaging principles, we uncover its capacity to capture intricate details and detect flaws without causing damage.The incorporation of machine learning enhances the capabilities of ultrasound-based non-destructive testing by enabling automated analysis, insights extraction, and improved decision-making. By highlighting the benefits and advancements in non-destructive testing, this presentation aims to inspire further exploration, collaboration, and innovation in leveraging ultrasound and machine learning for effective quality control and maintenance in industrial settings.
Call For Paper
Topics of the papers include but are not limited to:
- Computer Vision in Ultrasound Images;
- Algorithms to Mitigate Data Imbalance;
- Semi-supervised Learning in Ultrasound Images;
- Supervised Learning in Ultrasound Images;
- Multi-task Learning in Ultrasound Images;
- Deep Learning in Volumetric images;
- Data and Performance Baseline;
- Normalization Techniques in Ultrasound Image Analysis;
- Image Classification and Segmentation;
- Spatial-temporal Feature Analysis;
The selected papers will be presented as oral or poster presentations.
Each submitted paper should be limited to eight pages, including figures and tables , and must follow the same policies and submission guidelines described in CVPR'23 Author Guidelines (https://cvpr2023.thecvf.com/Conferences/2023/AuthorGuidelines).
Papers submission is through the CMT system. In submitting a manuscript to this workshop, the authors acknowledge that no paper with substantially similar content has been submitted to another workshop or conference during the review period.
See the demo notebook for more details of the metrics
Mean Square Distance
Residual Mean Distance