Semantic Segmentation Training Using Imperfect Annotations and Loss Masking

One of the most significant factors affecting supervised neural network training is the precision of the annotations. Moreover, the problem of inconsistent data annotations is an integral part of real-world supervised learning processes, concerning even annotations created by experts in the field. One practical example is a weak ground truth delineation for medical image segmentation. In this paper, we have developed a new method of accurate segmentation of blood vessels based on a convolutional neural network. We allowed training of the network with imperfect annotations for the semantic segmentation of blood vessels, and introduced a concept of uncertainty masks and loss masking. These uncertainty masks can be created roughly by non-experts, which makes annotation process cheaper and faster. We have tested our method quantitatively in two aspects: on a real-world blood vessel segmentation problem with missing vessel annotations, and on a perfectly labeled data set with artificially introduced annotation noise. Models trained with loss masking seem to be more robust regardless of the number of removed vessels. Noise robustness of four different model architectures has been tested and compared to the loss masking method, which was found to have superior performance when trained on noisy data.