@article{WANG2026, 
author = {Ziye WANG and Xiaohui TANG and Lan ZHOU and Chunyan XU and Shunping ZHOU and Kaiqiao ZHANG and Fangzhou LIU and Shengbin ZHOU},
title = {Segmentation of Fat Droplets in Whole Slide Images of Tissue Stained with Sudan Ⅲ},
year = {2026},
journal = {Forensic Science and Technology},
volume = {51},
number = {2},
pages = {121-128},
keywords = {forensic pathology, fat embolism, special staining, image segmentation, contrastive language-image pre-training (CLIP), deep learning, digital whole slide images},
url = {https://www.sciopen.com/article/10.16467/j.1008-3650.2025.0003},
doi = {10.16467/j.1008-3650.2025.0003},
abstract = {In forensic pathology, Sudan Ⅲ staining is used to confirm fat embolism, and its quantitative grading is of significant importance in determining the cause of death. However, manual grading based on microscopic observation is highly dependent on personal experience. To objectively quantify the degree of fat embolism, we explored a method for automatic segmentation of fat droplets in Whole Slide Images (WSI) of lung tissue stained with Sudan Ⅲ. Although the colors of the Sudan Ⅲ stained sections are simply consisted of transparent tissue and scarlet fat droplets, issues such as residual dye, uneven staining of the fat droplets, irregular shapes, and significant size differences can lead to missegmentation and insufficient segmentation accuracy. To address this, we propose a contrastive language-image pre-training (CLIP) model framework combined with prompt learning for fat droplet segmentation: first, feature maps output by the CLIP image encoder are fused through skip connections, guiding the model to accurately segment fat droplets using CLIP’s prior knowledge via text prompts; then, a dice loss function is used to alleviate the imbalance between the foreground and background of the image; finally, validation is performed on the slice dataset and compared with U-Net, FCN8s, and Unet++ models. The results indicate the method proposed in this article is superior to others in segmenting fat droplets on stained slice images. Moreover, the proposed cross-modal prompt learning can be integrated into other large segmentation models to perform specific target segmentation tasks.}
}