Task-Agnostic Noisy Label Detection via Standardized Loss Aggregation
2026-05-11 • Computer Vision and Pattern Recognition
Computer Vision and Pattern RecognitionArtificial Intelligence
AI summaryⓘ
The authors address the problem of incorrect labels in large medical imaging datasets, which can happen because different experts may disagree or some cases are unclear. They propose a method called Standardized Loss Aggregation (SLA) that measures how reliable each label is by looking at prediction errors across repeated validation tests. SLA gives a smooth score that reflects how likely a label is noisy, doing better than older counting methods. Their tests on eye images show SLA can quickly spot ambiguous or wrong labels, helping improve data quality for classification tasks.
noisy labelsmedical imagingcross-validationloss functionlabel reliabilityfold-level validationclassificationfundus datasetdata annotationperformance deviation
Authors
Inhyuk Park, Doohyun Park
Abstract
Noisy labels are common in large-scale medical imaging datasets due to inter-observer variability and ambiguous cases. We propose a statistically grounded and task-agnostic framework, Standardized Loss Aggregation (SLA), for detecting noisy labels at the sample level. SLA quantifies label reliability by aggregating standardized fold-level validation losses across repeated cross-validation runs. This formulation generalizes discrete hard-counting schemes into a continuous estimator that captures both the frequency and magnitude of performance deviations, yielding interpretable and statistically stable noisiness scores. Experiments on a public fundus dataset demonstrate that SLA consistently outperforms the hard-counting baseline across all noise levels and converges substantially faster, especially under low noise ratios where subtle loss variations are informative. Samples with high SLA scores indicate potentially ambiguous or mislabeled cases, guiding efficient re-annotation and improving dataset reliability for any classification task.