Weakly-Supervised Eczema Region Segmentation via Probabilistic Label Propagation and Semantic Boundary Prediction

Abstract In the field of weakly-supervised semantic segmentation, obtaining large-scale, densely labeled data for tasks such as eczema lesion segmentation remains a significant challenge. To address this, we propose a novel method that leverages sparse, cheaply acquired annotations to train a segmentation model effectively. By utilizing sparse image labelings as a starting point, our approach propagates these labels across the image to generate approximate dense labels. This label-propagation mechanism is formulated based on random-walk probabilities, which not only allows for efficient label inference but also provides uncertainty estimates that are integrated into the model’s loss function. We train a standard convolutional neural network (CNN) to predict the dense labels while jointly learning the label-propagation process. This strategy enables the model to capture fine-grained details of eczema regions, despite the absence of explicit supervision on boundary contours. Our experiments demonstrate that the proposed method achieves superior segmentation performance compared to conventional weakly-supervised approaches, illustrating its potential for real-world applications where labeled data is scarce.