Medical Image Segmentation with Noisy Labels
In medical field, it is expected to obtain more than one expert opinions to mitigate the possible mistakes in diagnoses. Similarly, pursuing the same goal, usually multiple annotations are collected for medical image analysis. Meanwhile, for training the supervised machine learning models the ground truth annotations are required and typically it is obtained from multiple raters via majority voting or one from the preferred expert or other label fusion approaches. However, this process losses important agreement/disagreement information from all experts. Our approach is motivated by coupled two convolutional neural network architectures: segmentation network that estimates the true annotation and annotation network that estimates pixel-wise confusion matrices for each rater for a given image. However, unlike others we want to make the first network to be confident in making the prediction of each class in the presence of the noisy labels alone. To accomplish it we introduce the regularization term that forces each pixel resulting from the first network to the identity matrix, with 1 being for the correct class and 0 for the rest. In addition, we make use of the fact there are regions where all annotators agree on and feed that information as input to the additional loss function. We conduct experiments on datasets with synthetic noises (on MNIST) and real-world noise (RIGA and KiTS).
A. Zhumabayeva, "Medical Image Segmentation with Noisy Labels", M.S. Thesis, Machine Learning, MBZUAI, Abu Dhabi, UAE, 2023.