A Comparative Study of Noisy Label Detection Techniques in a Thai Hospital's Chest X-Ray Database

Conference proceedings article

Authors/Editors

Strategic Research Themes

Publication Details

Author list: Tatitaisakul S., Wilainuch S., Chamveha I., Chaisangmongkon W.

Publisher: Institute of Electrical and Electronics Engineers Inc.

Publication year: 2024

Start page: 232

End page: 237

Number of pages: 6

ISBN: 9798350344349

URL: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85189939072&doi=10.1109%2fICAIIC60209.2024.10463316&partnerID=40&md5=c341303e9065856ade48c73d9839b1e3

Languages: English-Great Britain (EN-GB)

View on publisher site

Abstract

This paper addresses the problem of noisy labels in chest X-ray datasets, which significantly impact the training of deep neural network models. Noisy labels often occur due to errors in reports from experts or the use of algorithms to extract labels from medical reports written in natural language. To tackle this issue, we compared the effectiveness of O2U-net, a state-of-The-Art noisy label detection method, and NVUM, a noise-resistant model training technique in identifying noisy samples. We contrasted these methods with a heuristic approach which uses a simple classification model to flag samples with large differences between predicted and actual labels as noisy. Our findings indicated that NVUM outperformed the other methods in identifying noisy labels, providing a promising solution to the challenge of noisy labels in medical image analysis. © 2024 IEEE.

Keywords

Chest X-ray image, noise robust training, Noisy label detection