Structure Constrained Diffusion Models for Pterygium Segmentation

Pterygium is a progressive blinding disease, mostly encroaching on the cornea from the nasal bulbar conjunctiva. In clinical practice, pterygium is generally classified according to the degree of invasion into the cornea and pupil. Consequently, the detection and quantification of different ocular surface structure, i.e., pterygium, cornea, and pupil is of great importance for the accurate diagnosis and prediction of the disease progression. In this paper, we propose a Structure Constrained Diffusion Models (SCDM) for pterygium, cornea, and pupil segmentation in ocular surface images. The proposed method transforms pixel-level segmentation into a conditional generation task, which takes the ocular surface image as a condition to guide the diffusion models for generating the mask from Gaussian noise. To be more specific, we improve the model by using a multi-layer conditional guidance to reduce the randomness of diffusion models. We then develop a mixed loss function to increase the structure constraints of estimated mask in the reverse process. Finally, a novel deterministic sampling strategy is introduced to effectively accelerate the sampling process and provide the deterministic sampling space. Experimental results show that the proposed method outperforms other state-of-the-art methods by providing more accurate segmentation performances.