Caudo-Diff: Diffusion calibrated pseudo labels in guided latent space for minimally supervised medical segmentation

Accurate segmentation of medical images is essential for clinical diagnosis and treatment planning. However, deep learning-based segmentation models are data-intensive, requiring large, well-annotated datasets which is an often challenging and costly requirement in medical fields. To reduce the reliance on manual labeling, we propose the minimally supervision based on an exemplar, leveraging only a single labeled sample while making full use of the remaining unlabeled data. In this case, two challenges need to be addressed. First, a lack of sufficient prior information: relying solely on a single exemplar limits the model’s ability to capture complex semantics. Second, the unreliability of pseudo labels: noise and inaccuracies in these labels introduce bias, hindering segmentation performance. To overcome these challenges, we propose a new pseudo-labeling paradigm by diffusion calibration. Follow this paradigm, we introduce Caudo-Diff, a novel method for calibrating pseudo labels using a deterministic diffusion model in a guided latent space, aiming to supplement prior information and improve pseudo-labeling reliability. Initial pseudo labels and features extracted by the segmentation network guide the model to focus on meaningful semantic regions. The pseudo labels are then refined to reduce noise and errors, enhancing segmentation accuracy. Experimental results show that Caudo-Diff improves segmentation performance with minimal supervision, offering a practical solution to the challenge of annotation scarcity in medical image segmentation.