KIT | KIT-Bibliothek | Impressum | Datenschutz

What Your Radiologist Might be Missing: Using Machine Learning to Identify Mislabeled Instances of X-ray Images

Rädsch, Tim; Eckhardt, Sven; Leiser, Florian; Pandl, Konstantin D.; Thiebes, Scott; Sunyaev, Ali

Label quality is an important and common problem in contemporary supervised machine learning research. Mislabeled instances in a data set might not only impact the performance of machine learning models negatively but also make it more difficult to explain, and thus trust, the predictions of those models. While extant research has especially focused on the ex-ante improvement of label quality by proposing improvements to the labeling process, more recent research has started to investigate the use of machine learning-based approaches to identify mislabeled instances in training data sets automatically. In this study, we propose a two-staged pipeline for the automatic detection of potentially mislabeled instances in a large medical data set. Our results show that our pipeline successfully detects mislabeled instances, helping us to identify 7.4% of mislabeled instances of Cardiomegaly in the data set. With our research, we contribute to ongoing efforts regarding data quality in machine learning.

Open Access Logo

Postprint §
DOI: 10.5445/IR/1000127578
Frei zugänglich ab 06.01.2022
Zugehörige Institution(en) am KIT Institut für Angewandte Informatik und Formale Beschreibungsverfahren (AIFB)
Publikationstyp Proceedingsbeitrag
Publikationsdatum 05.01.2021
Sprache Englisch
Identifikator KITopen-ID: 1000127578
Erschienen in Proceedings of the 54th Hawaii International Conference on System Sciences (HICSS)
Veranstaltung 54th Hawaii International Conference on System Sciences (HICSS 2021), Online, 05.01.2021 – 08.01.2021
KIT – Die Forschungsuniversität in der Helmholtz-Gemeinschaft
KITopen Landing Page