Domain-specific data augmentation improves robustness of endoscopic AI systems
Endoscopy · 01.04.2024 · 56 (S 02 S 02) · S322-S323
DOI: 10.1055/s-0044-1783480 ↗Kurz gefasst
Fachartikel in der Zeitschrift Endoscopy, erschienen am 01.04.2024. Verfasst von M. Jong und 14 weiteren Autor:innen.
Abstract (Originalsprache)
Aims
The emergence of deep learning has significantly increased the amount of artificial intelligence (AI) systems in endoscopy. However, these systems are generally developed in expert centers with the use of uniform, high-quality imagery. In daily practice, imagery will be more heterogeneous, which will likely result in heavy degradation of AI performance. One aspect of this image heterogeneity is the use of different enhancement settings across endoscopy units. Endoscopy systems have numerous enhancement settings to improve perceived image quality. This study aims to quantify the impact of various enhancement settings on AI performance and proposes mitigation strategies.
Methods
In this study two commonly used AI systems in endoscopy were evaluated. A computer aided detection (CADe) system for Barrett’s neoplasia and a computer aided diagnosis (CADx) system for optical diagnosis of colorectal polyps. The images from the respective training and test sets were slightly augmented by adjusting post-processing enhancement settings. Performance variability was measured over all test sets with different enhancement settings. After initial performance assessment, both systems were retrained with all available enhancement settings.
Results
The CADe system for Barrett’s neoplasia reached AUC, sensitivity and specificity scores on the original test set of 96%, 89% and 89%. On test sets with adjusted enhancement settings, performance varied strongly with maximum deltas of 5%, 22% and 25% respectively. When trained on data augmented with all enhancement settings, this delta was reduced to 1%, 1% and 3%.The CADx system displayed AUC, sensitivity and specificity scores of 78%, 77% and 52%. Performance varied widely on test sets with altered enhancement settings with a maximum delta of 9%, 18% and 30%. After training on all enhancement settings, the CADx’ delta was 2%, 6% and 12%.
Conclusions
Performance of current existing AI systems may vary significantly depending on the selected post-processing enhancement setting of the deployment endoscopy unit. This variability can be effectively managed by domain-specific data augmentation, which ensures robust performance across different endoscopy centers. Publication History Article published online: 15 April 2024 © 2024. European Society of Gastrointestinal Endoscopy. All rights reserved. Georg Thieme Verlag KG Rüdigerstraße 14, 70469 Stuttgart, Germany
Zitieren
APA
Jong, M., Kusters, C., Bokhorst, Q. E. V., Jukema, J., Fockens, K., Heslinga, R. V. E. V., Jaspers, T., Boers, T., Houwen, B. B., Vlugt, M. V. D., Dekker, E., Sommen, F. V. D., With, P. D., Groof, J. D., & Bergman, J. (2024). Domain-specific data augmentation improves robustness of endoscopic AI systems. Endoscopy, 56(S 02 S 02), S322-S323. https://doi.org/10.1055/s-0044-1783480
Vancouver
Jong M, Kusters C, Bokhorst QEV, Jukema J, Fockens K, Heslinga RVEV, et al. Domain-specific data augmentation improves robustness of endoscopic AI systems. Endoscopy. 2024;56(S 02 S 02):S322-S323. doi:10.1055/s-0044-1783480
BibTeX
@article{Jong2024,
author = {Jong, M. and Kusters, C. and Bokhorst, Q.N. E. van and Jukema, J. and Fockens, K. and Heslinga, R. van Eijck van and Jaspers, T. and Boers, T. and Houwen, B. B. and Vlugt, M. Van Der and Dekker, E. and Sommen, F. Van Der and With, P. De and Groof, J. De and Bergman, J.},
title = {Domain-specific data augmentation improves robustness of endoscopic AI systems},
journal = {Endoscopy},
year = {2024},
volume = {56},
number = {S 02 S 02},
pages = {S322--S323},
doi = {10.1055/s-0044-1783480}
}
EUVIMED