New LiNC Method Corrects Label Noise in Medical Imaging Datasets

Label noise is a significant challenge in medical imaging datasets, often stemming from human variability or annotation errors, which can compromise the reliability of machine learning models. A new technique, Lightweight Noise Correction (LiNC), addresses this by integrating a unique trainable "trust" parameter for each training sample. This parameter helps the model decide whether to rely on the observed label or its own prediction during training. LiNC operates in phases: initially, it uses the gradient of its objective function to differentiate between clean and noisy samples, creating distinct trust distributions. A 3-component Gaussian Mixture Model then categorizes samples into clean, ambiguous, or noisy. Subsequently, the method applies a soft correction to noisy cases, followed by a hard correction phase, leading to improved label quality. Evaluations on ten 2D MedMNISTv2 datasets, even with up to 50% label noise, demonstrated consistent accuracy gains and effective mislabel detection. The method is computationally efficient, adding only negligible overhead to the base network's training time and memory footprint.