Show abstract - PDF - Reviews

Smoothing hard-label assignments has emerged as a popular strategy in training discriminative models. Nevertheless, most existing approaches are typically designed for classification tasks, ignoring underlying properties of dense prediction problems, such as medical image segmentation. First, these strategies often ignore the spatial relations between a given pixel and its neighbours. And second, the image context associated with each label is overlooked, which can convey important information about potential errors or ambiguities in the segmentation masks. To address these limitations, we propose in this work geodesic label smoothing (GeoLS), which integrates image information into the label smoothing process by leveraging the geodesic distance transform of the images. As the resulting label assignment is based on the computed geodesic map, class-wise relationships in the soft-labels are better modeled, as it considers image gradients at the boundary of two or more categories. Furthermore, spatial pixel-wise relationships are captured in the geodesic distance transform, integrating richer information than resorting to the Euclidean distance between pixels. We evaluate our method on two publicly available segmentation benchmarks and compare them to popular segmentation loss functions that directly modify the standard hard-label assignments. The proposed geodesic label smoothing improves the segmentation accuracy over existing soft-labeling strategies, demonstrating the validity of integrating image information into the label smoothing process. The code to reproduce our results is available at: https://github.com/anonymous35783578/GeoLS.
Hide abstract