Scribble Supervised Segmentation

Scribble-supervised segmentation aims to train accurate image segmentation models using only sparse user-provided scribbles instead of laborious pixel-level annotations. Current research focuses on improving model architectures, such as incorporating Transformers alongside Convolutional Neural Networks, and developing novel training strategies like model mixup and generative data augmentation to overcome the limitations of weak supervision. This approach significantly reduces annotation effort, making high-quality segmentation accessible for applications where fully labeled data is scarce, particularly in medical image analysis. The resulting advancements are driving progress in label-efficient learning and expanding the applicability of deep learning to diverse domains.