Influence Based Re-Weighing for Labeling Noise in Medical Imaging

Labeling for pathology detection and grading in medical imaging is laborious and error-prone. High inter-rater variability is common even among board-certified sub-specialty radiologists. A typical approach to identify and correct labeling errors is to have experts with the same qualification review all the rated cases again, which is inevitably very costly. To reduce the expert annotation effort, we propose a re-weighing method based on TracIn [1], a labeling noise identification method, to assign low weights to potentially noisy samples during training. The weights are calculated based on self-influence scores from TracIn. The re-weighing TracIn leads to significantly improved performance for lumbar spine stenosis detection in MRI, with the macro accuracy of 77.2%, compared against the baseline accuracy of 75.4%, and the TracIn without re-weighing accuracy of 76.6%. The proposed re-weighting TracIn is also agnostic to the choice of training models, and thus can generalize to to other deep learning approaches. Also, to increase data efficiency, we investigated unsupervised data augmentation (UDA) and contrastive self-supervised learning (SimCLR) to utilize unlabeled data.