GLANCE: Combating Label Noise using Global and Local Noise Correction for Multi-Label Chest X-ray Classification

04 Aug 2024 (modified: 01 Sept 2024)MICCAI 2024 Workshop UNSURE SubmissionEveryoneRevisionsBibTeXCC BY 4.0
Keywords: Chest X-Ray classification, label noise, label co-occurrence
Abstract: Chest X-ray imaging is essential for diagnosing thoracic dis eases, with multi-label classification playing a critical role in identifyingmultiple conditions from a single image. Despite deep neural networks(DNNs) significantly advancing this field, the reliance on noisy labels ex tracted from clinical reports poses a significant challenge, underminingDNNs’ performance. Several research attempt to address this issue butfail to consider the critical inter-class correlations prevalent in CXR diag nostics. To this end, we propose a Global and LocAl Noise CorrEction(GLANCE) framework. The GLANCE framework comprises a classifi cation backbone and two primary components: a global noise correc tion (GNC) module and a local noise correction (LNC) module. TheGNC module calculates the noise transition matrix based on the labelco-occurrence frequencies and uses the transition matrix to reduce theimpact of the noisy labels. The LNC module treats the temporal en sembling of samples historical predictions as the instance-specific pseudolabels, which also serve as the supervision. Our GLANCE framework addresses the shortcomings of existing techniques, i.e., the unreliability ofnoise transition matrices in the presence of class imbalances and zero co occurrence frequencies. Comprehensive experimental results demonstrate that GLANCE surpasses competing methods, showcasing its superiorability to combat label noise and improve multi-label CXR classification accuracy.
Submission Number: 18
Loading