Efficient Transformers for X-Ray NDT in limited data scenarios
e-Journal of Nondestructive Testing · 2026-07 · 31 (7)
DOI: 10.58286/33444 ↗At a glance
Journal article in e-Journal of Nondestructive Testing, 2026-07. Written by Ihssane Ghalas and 1 co-authors.
Abstract (original language)
Non-destructive testing (NDT) increasingly leverages deep learning, but the lack of large annotated X-ray datasets limits model deployment. Vision Transformers (ViTs) excel across vision tasks but require substantial training data, making direct application to X-ray NDT difficult. Self-supervised learning (SSL) enables ViTs to learn transferable features without labels, though pretraining on natural images like ImageNet is limited by domain gaps. To better align with industrial needs, we pretrain on medical X-ray datasets and investigate parameter efficient fine-tuning (PEFT) methods. Results show that domain-relevant SSL pretraining combined with PEFT strategies greatly enhances performance in limited-data X-ray NDT scenarios, providing an effective solution for industrial inspection.
Cite
APA
Ghalas, I., & Vandoni, J. (2026). Efficient Transformers for X-Ray NDT in limited data scenarios. e-Journal of Nondestructive Testing, 31(7). https://doi.org/10.58286/33444
Vancouver
Ghalas I, Vandoni J. Efficient Transformers for X-Ray NDT in limited data scenarios. e-Journal of Nondestructive Testing. 2026;31(7). doi:10.58286/33444
BibTeX
@article{Ghalas2026,
author = {Ghalas, Ihssane and Vandoni, Jennifer},
title = {Efficient Transformers for X-Ray NDT in limited data scenarios},
journal = {e-Journal of Nondestructive Testing},
year = {2026},
volume = {31},
number = {7},
doi = {10.58286/33444}
}
EUVIMED