Diffusion Model in Semi-Supervised Scleral Segmentation

Accurate scleral segmentation is crucial in the visual diagnosis of Traditional Chinese Medicine. However, acquiring annotated labels for training such images is labor-intensive and raises privacy concerns during the annotation process. To address this issue, semi-supervised learning (SSL) can effectively reduce annotation costs. Nevertheless, the presence of a large number of unlabeled data images poses a challenge to semi-supervised learning. This paper introduces a semi-supervised two-stage automatic data enhancement framework. Initially, an extended scattering model is introduced to enhance the data in the first stage. In the second stage, two semi-supervised learning frameworks are experimented with: one utilizing domain-specific enhancement and the other employing a semi-supervised basis with spatial variation and an inverse permutation framework. Furthermore, a TCM visual diagnosis dataset was collected in our laboratory, and multiple ablation experiments were conducted on this framework to demonstrate the substantial impact of the introduced diffusion model. The experimental results demonstrate that the introduction of the diffusion model enables excellent segmentation with a minimal amount of data. Moreover, accurate scleral segmentation within the comprehensive TCM ophthalmology dialectic framework establishes the foundation for extracting pathological vessel filaments.