LiNC: A Lightweight Method for Correcting Label Noise in Medical Imaging Datasets
A new approach for addressing label noise in medical imaging datasets has been shared in a preprint on arXiv. This noise, which can arise from things like differences in raters, mistakes in labeling, and unclear cases, poses a risk to the effectiveness of machine learning models in clinical settings. The technique is called Lightweight Noise Correction (LiNC) and introduces a single adjustable trust parameter for each training example. It helps the model figure out when to trust the given label versus its own predictions during training. By blending the actual label with the model's predictions, it fine-tunes trust levels for both clean and noisy data, using a 3-component Gaussian Mixture Model for correction. You can check out the paper on arXiv under the ID 2608.04147.
Key facts
- Label noise is common in medical imaging datasets due to inter-rater variability, annotation errors, and ambiguous cases.
- LiNC adds a single trainable trust parameter per training sample.
- The method learns when to use the observed label and when to defer to the model during a standard training loop.
- Training uses a convex combination of the observed label and the model's predictive distribution, controlled by the per-sample trust parameter.
- The gradient of the objective drives trust values in opposite directions for clean versus noisy samples in the early training phase.
- A 3-component Gaussian Mixture Model is used to separate the trust distributions.
- The paper is available on arXiv with identifier 2608.04147.
- The method aims to improve reliability and clinical effectiveness of machine learning models trained on noisy datasets.
Entities
Institutions
- arXiv