Stochastic co-teaching for robust cardiac segmentation in ultrasound with noisy labels

In this work, we propose a label noise-robust segmentation framework for left ventricle blood pool segmentation in echocardiography. Based on the stochastic co-teaching approach, our method extends pixel-level filtering of label noise with additional image-level filtering to more effectively prevent noisy labels from backpropagating. We evaluate our framework on the EchoNet-Dynamic dataset, and simulate diverse noisy label scenarios, including over- and undersegmented (i.e., biased) labels. Our results demonstrate that the incorporation of image-based rejection enhances the Dice coefficient by 1.5% points and ejection fraction estimation by 2.3% points with respect to the pixel-based co-teaching framework under heavily biased label noise conditions, and thereby maintains the same performance as on clean data.