Spatial Reasoning Loss for Weakly Supervised Segmentation of Skin Histological Images

Biological images often follow some kind of geometrical structure. For example, histological images of reconstructed human skin follow a structural order with the stratum corneum as the outer layer and the living epidermis just below it. In this paper such spatial relationships are leveraged to define a loss function that penalizes structures that do not respect the given pattern as a form of weak supervision. The proposed loss function is based on fuzzy ontological spatial reasoning and morphological operators. The model is tested in a segmentation task on skin images, where a small number of labeled images and a large number of unlabeled ones are available. The proposed method leverages information in unlabeled images to improve segmentation results, compared with training only on the labeled data.
¹