Weakly Supervised Segmentation of Subcutaneous Blood Vessels using Semantic Boundary Propagation and Uncertainty Estimation

Abstract We address the challenge of training deep neural networks for subcutaneous vascular segmentation with limited labeled data by leveraging weak supervision. Given sparse and inexpensive annotations, we introduce a label propagation strategy that generates dense pseudo-labels from the sparse inputs. To train a convolutional neural network (CNN) for segmentation, we align the network's predictions with these propagated labels. The label propagation is formulated through a random walk framework, utilizing hitting probabilities to model the spread of labels across the image, while incorporating uncertainty estimates into the loss function. This approach allows for the joint optimization of the label propagation mechanism and the segmentation model, enabling the network to learn fine-grained vascular boundaries without requiring explicit edge-level supervision. Our results demonstrate that this weakly supervised methodology significantly improves segmentation accuracy compared to conventional methods trained with sparse annotations.