CPS4: Class Prompt driven Semi-Supervised Spine Segmentation with Class-specific Consistency Constraint
Researchers have introduced CPS4, a novel text-guided semi-supervised network for spine segmentation. This method leverages Vision Language Models (VLMs) with class prompts to improve the quality of pseudo-labels in segmentation tasks. CPS4 employs a two-stage training process, first optimizing consistency between textual prompts and image regions, then using the pretrained encoder to generate class-specific segmentation maps for unlabeled data. The approach achieved a Dice score of 80.44% using only 5% labeled data on a public dataset, outperforming existing semi-supervised and VLM methods. AI
IMPACT Introduces a novel approach for improving segmentation accuracy in medical imaging using VLMs and class prompts.