Reproduce “Active Label Refinement for Robust Training of Imbalanced Medical Image Classification Tasks in the Presence of High Label Noise” on OpenResearch

About this paper

Bidur Khanal, Tianhong Dai, Binod Bhattarai, Cristian Linte

A method combining Co-teaching with Variance of Gradients (CT_VOG) and active learning enables robust training of deep learning models on highly imbalanced medical image datasets afflicted by severe label noise. This approach effectively identifies and leverages clean samples, leading to performance comparable to training on fully clean data by relabeling only a small fraction of the dataset.

Key insights

• Incorporating Variance of Gradients (VOG) into noisy label learning methods helps accurately distinguish truly noisy labels from challenging, clean minority class samples in imbalanced datasets.
• Initiating model training on a robustly identified clean subset, rather than fine-tuning a model pre-trained on noisy data, improves final performance.
• Combining an initial robust learning phase for noisy, imbalanced data with subsequent active label cleaning is more effective than either approach in isolation.

Summary by alphaXiv.