To enhance the efficiency of inspections along with the system digitalization through DDA and CR, we propose three novel solutions with artificial intelligence for NDT.
First, “Workflow AI” assists inspectors by automatically determining whether the required quality of image is realized with appropriately placed IQIs. The AI not only detects IQIs but also analyzes them, annotating the DICOM image with statistical information, which used to be manually measured.
Second, “Wall-Thickness AI” enables automatic calculation of pipe wall thickness based on pixel-wise pipe boundary data derived by a background AI, allowing inspectors to instantly identify the most severely corroded point within arbitrary area of interest. Compared to existing calculation tools, boundary determination accuracy for irregular shapes has improved due to our study on profile analysis of both X-ray images and our partner inspectors’ judgement data.
Third, “Screening AI” automatically highlights defects and possibly unacceptable indications in radiographic images for welding or casting. It demonstrates consistent true positive rates of over 90% for indications like porosity, gas hole and slag inclusion, though it is not limited to these types. Attribute information for the detected indications, including their actual size and their total numbers, can be displayed in a list, supporting inspectors to easily judge the grade of defects.
In developing these three solutions, we placed value on on-site data collected through collaboration with customers, which not only contributes directly as training data to improve AI performance but is also utilized to verify the AI effectiveness. Case studies demonstrated that the proposed applications lead to faster inspections and enhanced stable quality in NDT daily operations.

This paper deals with the implementation of Aritificial Intelligence and supervised Machine Learning to train Large Language Models to read and analyze 8bit monochromatic images and provide basis for autonomous radiographic interpretations.
The setup involved analysis of approximately 200 digital images, interpreted by three independent ASNT Level 3 Professionals and these results are loaded as foundation models which learn about the patterns through this training data. This is made possible by employing a Conditional Probability Distribution that Discriminates between kinds of data instances and classifies defects as trained by the initial data.
These trained models are formed as a database which can either be used as a generative AI or for further ‘Deep Learning’ of these models that uses Neural Networks allowing them to process more complex patterns than traditional machine learning.

The trained model was then allowed to interpret another 50 images, which had a surprising accuracy of 82%. There were misinterpretations with regards to film artefacts, and spurious indications which a regular interpreter would easily identify. Given this success, with more training data and digital images, the model can be trained to create a self-sustainable and robust autonomous interpretation engine.

Gas utilities are under increasing pressure to reduce failure risk, standardize inspection quality, and modernize legacy radiographic records. This paper presents results from an AI-assisted radiography program using digitized pipeline weld film, focused on indication flagging, sizing, and segmentation, and on quantifying variance between human interpreters and multiple AI tools.

A curated set of pipeline girth weld radiographs from multiple sources was digitized and reviewed by qualified interpreters using structured Gage R&R methods to establish ground truth for indication presence, length/height sizing, and region-of-interest segmentation. Particular emphasis was placed on high-variance indication types (e.g., incomplete fusion, internal concavity, and related weld profile indications) and on films of differing image quality and exposure conditions.
Building on this foundation, OOGA implemented a streamlined ADR workflow that integrated image ingestion, human review, AI evaluation, and result reporting in a single environment. Several commercial and open-source AI “lanes” were then applied to the same dataset. Performance was evaluated in terms of indication flagging (probability of detection and false-call behavior), sizing variance relative to the human ground truth, and consistency of segmentation across indication types and film quality levels. As a result, some AI partners were able to evolve from proof-of-concept tools to fully functional modules aligned with practical operator workflows, including additional quality verification steps that mirror existing NDT practices.

Results show that while operator variance can be substantial for certain indication classes and lower-quality films, AI-assisted analysis can provide more consistent flagging and sizing on well-digitized images, with performance that is comparable to or better than the average human reader in specific indication categories. The paper will present variance trends, comparative AI vs. operator performance by indication type and film quality, and a roadmap for embedding AI-assisted ADR and verification steps into utility procedures, training, and quality assurance workflows.

Three-dimensional X-ray microscopy (XRM) has emerged as a transformative tool for non-destructive characterization of complex materials, enabling multiscale imaging and quantitative analysis of microstructural features. However, conventional reconstruction algorithms impose trade-offs between resolution, field of view, and acquisition time, limiting throughput and statistical representativity for large-volume studies.
In this work, we present an AI-driven reconstruction framework that integrates deep learning algorithms—such as ZEISS DeepRecon Pro and Advanced Reconstruction Toolbox (ART)—to overcome these limitations. By leveraging convolutional neural networks our approach achieves significant improvements in image quality, artifact suppression, and resolution recovery without increasing acquisition time. Experimental validation on nickel-based metal matrix composites reinforced with TiC demonstrates up to 10× throughput enhancement and sub-micron resolution recovery in dense materials, compared to traditional Feldkamp-Davis-Kress (FDK) methods.
The proposed workflow enables accurate segmentation of secondary phases, porosity quantification, and defect detection across large volumes, facilitating correlative studies and accelerating materials development. These advancements position AI-enabled XRM reconstruction as a critical enabler for high-throughput characterization in energy storage, additive manufacturing, and advanced alloy systems.