Tang, Yucheng.; He, Yufan.; Nath, Vishwesh.; Guo, Pengfeig.; Deng, Ruining.; Yao, Tianyuan.; Liu, Quan.; Cui, Can.; Yang, Yuechen.; Yin, Mengmeng.; Xu, Ziyue.; Roth, Holger.; Xu, Daguang.; Yang, Haichun.; Huo, Yuankai. (2026).听.听IS and T International Symposium on Electronic Imaging Science and Technology, 38(11), 1991鈥�1996.听
Traditional deep learning methods for segmenting medical images in digital pathology usually work in two steps: they first split a very large whole slide image into small pieces, called patches, and then stitch the results back together. This can miss important fine details and broader tissue context, especially for whole slide images, which are so large that they can contain more than 80,000 by 70,000 pixels. In this paper, we introduce HoloHisto, a new method designed to segment these extremely large pathology images end to end, meaning the system can analyze the full image directly and produce the matching segmentation mask without relying on a patch-and-rebuild approach. HoloHisto uses a large 4K starting patch to capture more visual information at once and a new sequential tokenization step, which converts image features into a structured set of smaller units so the model can better understand relationships across the image while processing it efficiently. To our knowledge, this is the first holistic method for segmenting gigapixel whole slide images and can handle both the full image and its corresponding pixel-level mask directly. We also introduce a random 4K sampling strategy that provides far more image information than standard smaller patches. To test the method, we created a new kidney pathology dataset with whole-slide segmentation of glomeruli, the tiny filtering structures in the kidney, from entire mouse kidneys. The results show that HoloHisto-4K performs substantially better than previous state-of-the-art methods.
