Confidence-Guided Learning For Breast Density Classification
Sun Young Park, Dustin Sargent, David Richmond · 2021
Assessment of breast density is an important part of breast cancer screening. Dense tissue is a risk factor and increases the chance of findings being occluded and missed by a radiologist. Density is an inherently subjective assessment, and the inter-rater variability between radiologists has been reported at 30-40%. The subjectivity of breast density assessment means that there is no “gold standard” label for training an algorithm. To address this issue, we created a multi-annotator dataset, where each case is labeled multiple times independently, along with annotator confidence. To utilize this information during algorithm training, we introduce a confidence-guided method that adjusts batch distribution, batch size and loss calculations based on consistency and confidence of the labels. Using this approach, we achieve breast density classification that is interchangeable with the radiologist annotations on a test set from an independent held-out site, simulating a deployment of the algorithm in a real-world scenario.