Impact of Noisy Labels on Dental Deep Learning—Calculus Detection on Bitewing Radiographs
Journal of Clinical Medicine · 23.04.2023 · 12 (9) · 3058
DOI: 10.3390/jcm12093058 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Clinical Medicine, online erschienen am 23.04.2023. Verfasst von Martha Büttner und 5 weiteren Autor:innen. Beteiligte Einrichtung: Department of Oral Diagnostics, Digital Health and Health Services Research, Charité-Universitätsmedizin Berlin, 14197 Berlin, Germany. Die Arbeit wurde bisher 17-mal zitiert (Crossref) und nennt 21 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: World Health Organization, Charité—Universitätsmedizin Berlin, German Research Foundation (DFG).
Abstract (Originalsprache)
Supervised deep learning requires labelled data. On medical images, data is often labelled inconsistently (e.g., too large) with varying accuracies. We aimed to assess the impact of such label noise on dental calculus detection on bitewing radiographs. On 2584 bitewings calculus was accurately labeled using bounding boxes (BBs) and artificially increased and decreased stepwise, resulting in 30 consistently and 9 inconsistently noisy datasets. An object detection network (YOLOv5) was trained on each dataset and evaluated on noisy and accurate test data. Training on accurately labeled data yielded an mAP50: 0.77 (SD: 0.01). When trained on consistently too small BBs model performance significantly decreased on accurate and noisy test data. Model performance trained on consistently too large BBs decreased immediately on accurate test data (e.g., 200% BBs: mAP50: 0.24; SD: 0.05; p < 0.05), but only after drastically increasing BBs on noisy test data (e.g., 70,000%: mAP50: 0.75; SD: 0.01; p < 0.05). Models trained on inconsistent BB sizes showed a significant decrease of performance when deviating 20% or more from the original when tested on noisy data (mAP50: 0.74; SD: 0.02; p < 0.05), or 30% or more when tested on accurate data (mAP50: 0.76; SD: 0.01; p < 0.05). In conclusion, accurate predictions need accurate labeled data in the training process. Testing on noisy data may disguise the effects of noisy training data. Researchers should be aware of the relevance of accurately annotated data, especially when testing model performances.
Schlagwörter
Stichwörter der Autor:innen
Autor:innen und Institutionen
- Department of Oral Diagnostics, Digital Health and Health Services Research, Charité-Universitätsmedizin Berlin, 14197 Berlin, Germany.
- ITU/WHO Focus Group AI4Health, Topic Group Dental Diagnostics and Digital Dentistry, CH-1211 Geneva 20, Switzerland.
ORCID-Profile: Martha Büttner, Aleksander Krasowski, Ben Feldberg, Falk Schwendicke
Förderung
- World Health Organization
- Charité—Universitätsmedizin Berlin
- German Research Foundation (DFG)
Zitieren
APA
Büttner, M., Schneider, L., Krasowski, A., Krois, J., Feldberg, B., & Schwendicke, F. (2023). Impact of Noisy Labels on Dental Deep Learning—Calculus Detection on Bitewing Radiographs. Journal of Clinical Medicine, 12(9), 3058. https://doi.org/10.3390/jcm12093058
Vancouver
Büttner M, Schneider L, Krasowski A, Krois J, Feldberg B, Schwendicke F. Impact of Noisy Labels on Dental Deep Learning—Calculus Detection on Bitewing Radiographs. Journal of Clinical Medicine. 2023;12(9):3058. doi:10.3390/jcm12093058
BibTeX
@article{Bttner2023,
author = {Büttner, Martha and Schneider, Lisa and Krasowski, Aleksander and Krois, Joachim and Feldberg, Ben and Schwendicke, Falk},
title = {Impact of Noisy Labels on Dental Deep Learning—Calculus Detection on Bitewing Radiographs},
journal = {Journal of Clinical Medicine},
year = {2023},
volume = {12},
number = {9},
pages = {3058},
doi = {10.3390/jcm12093058}
}
EUVIMED