Task-Agnostic Noisy Label Detection via Standardized Loss Aggregation

Researchers Inhyuk Park and Doohyun Park have proposed a new framework called Standardized Loss Aggregation (SLA) to address the prevalent issue of noisy labels in large-scale medical imaging datasets. These errors often arise from inter-observer variability and ambiguous cases. The SLA framework is statistically grounded and task-agnostic, designed to detect noisy labels at the sample level by aggregating standardized fold-level validation losses across repeated cross-validation runs. Unlike traditional discrete hard-counting schemes, SLA functions as a continuous estimator that captures both the frequency and magnitude of performance deviations, providing interpretable and stable noisiness scores. Experimental results on a public fundus dataset demonstrate that SLA consistently outperforms baseline methods across various noise levels and converges significantly faster, particularly in low-noise scenarios where subtle loss variations are critical. High SLA scores effectively identify potentially mislabeled or ambiguous cases, facilitating efficient re-annotation processes. This approach aims to enhance dataset reliability for any classification task, offering a robust solution for improving the quality of training data in computer vision and medical AI applications.