Short answer
Industrial X-ray and CT inspection images are typically held by foundries, weld shops and non-destructive testing providers that already radiograph parts. Their value depends on indication labels, the acceptance standard applied, exposure metadata and links to disposition. Rights are often shared with the part owner. The images can support internal-defect detection models, not robot manipulation or control.
The buyer's problem
Visible-light cameras see surfaces. Porosity inside a cast housing, a lack of fusion buried in a weld or a void in a bonded joint does not reach the surface, so models for those defects typically need radiographic or tomographic images. Public material in this modality is thin and narrow, which sends buyers to the organizations that generate the images in routine work: foundries that radiograph castings, fabricators that film or digitally image welds, and independent NDT service companies.
The images are only half of what makes the archive usable. A radiograph without the reader's call, the acceptance criterion and the technique that produced it is a grayscale picture that you have to relabel from scratch, and relabeling X-ray images typically calls for qualified readers.
A public reference, and where it stops
GDXray collects 19,407 X-ray images in five groups: castings, welds, baggage, natural objects and settings, with annotations as bounding boxes or labels (Mery et al., 2015). It is free for research and education, which is a licensing scope, not a commercial one.
Use it to see what a labeled NDT set can look like and to prototype pipelines. Do not treat its castings and welds groups as a proxy for your parts: different alloys, wall thicknesses, detector types and defect populations produce different images. Its licensing scope also means it cannot simply be mixed into a commercial training set without checking the terms; this is not legal advice, and your counsel should read the terms that apply to your use.
Radiography versus CT in a purchase
The two modalities produce different data objects, and a requirement should name which it wants.
| Aspect | Radiograph (2D) | CT (volume) |
|---|---|---|
| Data object | One projection image per exposure or shot | Reconstructed volume, often many slices, sometimes with raw projections |
| Typical label | Indication box or mask per image | Region or voxel mask; per-slice labels are common but incomplete |
| Metadata that matters | Source energy, exposure, detector, part orientation | Reconstruction settings, voxel size, scan geometry |
| Storage and delivery | Moderate; image files | Large; delivery logistics need planning |
| Typical supplier | Foundries, weld inspection contractors | Labs and manufacturers with CT capacity |
Castings and welds are frequent subjects; see casting and molding defect data and weld inspection data for process-specific detail. Volumetric data raises delivery questions covered by the general defect annotation formats and delivery page. For 3D surface rather than internal data, see 3D surface inspection data.
What makes an NDT archive worth licensing
Indication labels with a taxonomy. A reader's note reading "reject" is not enough. Ask for indication types, such as gas porosity, shrinkage, inclusions, cracks or incomplete fusion, and locations. A common taxonomy across years matters more than a long one; see defect taxonomy and label schema.
The acceptance standard behind the call. The same indication can pass under one specification and fail under another, and a customer specification may be stricter than a published standard. This page cites no radiographic acceptance standard, so ask the supplier to name the documents it applied and read them directly. A label is only meaningful together with the criterion that produced it; acceptance standards as label definitions explains the consequences.
Technique records. Exposure parameters, geometry, image quality indicators and detector type determine which defects could be seen at all. Missing records limit what the model can claim. Asking the supplier to document acquisition in the style of Datasheets for Datasets is one way to capture this.
Reader identity and review level. Ask whether one qualified reader or a second reader made each call, and whether disputed images were escalated. Agreement between readers is a quality signal; see inter-inspector agreement.
Link to outcome. A rejected casting that was repaired and reshot is a different record from one that was scrapped. Part and heat or lot identifiers allow joining to disposition.
Rights, which are more tangled than usual
NDT providers often image parts that belong to their customers, under service contracts that may restrict reuse of results. Foundries may hold images of customer-owned designs. Parts for regulated, aerospace or defense-related programs can carry additional restrictions. Questions to put to any supplier: who owns the part design, who owns the images, whether the service contract allows the supplier to license images derived from a customer's parts, and whether an end customer must consent. This is not legal advice. Blanking part numbers on an image header does not settle ownership.
Radiation safety is a separate diligence item. X-ray and CT capture uses ionizing radiation, so any commissioned capture would need to run under the provider's own radiation-safety program and applicable licensing, not be improvised for a data project. Ask how existing images were produced and who would perform any new capture; this page is not safety guidance.
Where a supplier is an inspection firm, third-party inspection firms as suppliers goes into how that changes diligence, and confidentiality in defect data covers what leaks through images.
Illustrative package: radiographic casting set
# Illustrative only - synthetic fields and values, not an available dataset
package: radiograph_casting_v0
supplier_type: foundry_ndt_lab
images:
format: 16-bit grayscale TIFF, original dynamic range kept
count_basis: "buyer sets; depends on defect population needed"
per_image_fields:
part_family: housing_type_X
lot_id: lot_0831
exposure_note: "energy and exposure recorded in technique sheet T-12"
view_number: 3
reader_id: reader_B
second_reader: true
indications:
- {type: gas_porosity, box: [x, y, w, h]}
acceptance_ref: "customer spec rev C, class 2"
disposition: accept | repair | scrap
excluded: robot actions, joint states, force data
rights: "training and evaluation; part owner consent recorded"
What this data cannot support
X-ray and CT images describe the internal state of a part at inspection time. They contain no manipulation, so they cannot train a humanoid to handle, place or load parts; control learning needs robot or teleoperation data, and robot manipulation failure datasets are a separate topic. They also cannot show what the defect did later in service unless joined to warranty and field failure data.
They are poorly suited to anomaly detection from normal-only images if the archive contains only rejects, and class imbalance is usual: rejects tend to be few. Detector and source differences make cross-site transfer uncertain, so hold out by plant, detector or technique when the claim is generalization. As ever, the buyer's own validation on its parts is what counts.
Questions to ask suppliers
- Which standards, customer specifications or internal criteria were used to accept or reject?
- Are exposure, detector and geometry parameters stored per image or per shot?
- Who read the images, with what qualification, and how often was a second reader used?
- Are images stored in their original bit depth, or exported as compressed previews?
- For CT, are raw projections retained or only reconstructed volumes?
- Can images be joined to part, heat or lot IDs, and to repair and scrap outcomes?
- What does the service contract say about reuse of results from customer parts?
- Are any images from export-controlled or restricted programs, and how are they flagged?
- How are archived films or files digitized, and was anything lost?
Discuss your requirements
If you need radiography or CT inspection data, describe the part type, defect classes, modality and permitted uses to SourceX. SourceX looks for US businesses that may hold suitable material, checks the data and the supplier's licensing permissions, and follows up about potential sources; a request does not guarantee a match. Start from the inspection data hub or the wider humanoid training data guide, and see the catalog entry for manufacturing quality records.
Explore sourcing and licensing options with SourceX
Sources
- GDXray: The Database of X-ray Images for Nondestructive Testing — public NDT X-ray database, its image groups, annotation form and research-and-education licensing scope.
- Datasheets for Datasets — documentation practice for recording how a dataset was acquired.