MSMTSeg: Multi-Stained Multi-Tissue Segmentation of Kidney Histology Images via Generative Self-Supervised Meta-Learning Framework
Xueyu Liu, Rui Wang, Yexin Lai, Yongfei Wu, Hangbei Cheng, Yuanyue Lu, Jianan Zhang, Ning Hao, Chenglong Ban, Yanru Wang, Shuqin Tang, Yuxuan Yang, Ming Li, Xiaoshuang Zhou and Wen Zheng
We present a generative self-supervised meta-learning framework to implement multi-stained multi-tissue segmentation, namely MSMTSeg, from renal biopsy WSIs using only a few annotated samples for each stain domain. MSMTSeg consists of multiple stain transform models for achieving inter-translation between multiple stain domains, a self-supervision module to obtain a pre-trained weight with individual information for each stain, and a meta-learning strategy combined with generated virtual data and the pre-trained weights to learn the common information across multiple stain domains and improve segmentation performance.
The complete code will be made public later
Due to medical ethics issues, our data will be conditionally disclosed after a public ethical review. Interested parties will request permission to access the data by emailing the address: liuxueyu0229@link.tyut.edu.cn.
Thanks Cycle-Gan, MoCov3, MAML, Unet. for serving as building blocks of MSMTSeg.