Skoltech等团队综述:五种AI工具在肺癌、乳腺癌、结直肠癌和前列腺癌组织分析中的表现对比
Researchers compare AI tools for analyzing tumor tissue
来自俄罗斯斯科尔科沃理工学院、阿联酋沙迦大学和日本东海大学的团队在《Discover Artificial Intelligence》发表开放获取综述,系统比较Cellpose、StarDist、CellProfiler、Icy和Ilastik五种计算机视觉工具在肺、乳腺、结直肠和前列腺癌组织图像识别上的表现。
Researchers at Skoltech, the University of Sharjah in the United Arab Emirates and Tokai University in Japan, together with colleagues, have published an open-access review of computer vision in histopathology. The review examines how algorithms can identify tissue features in lung, breast, colorectal and prostate cancers.
These findings will help medical professionals and developers choose the most suitable tools for specific clinical tasks, making image analysis more accurate and reproducible. The article appears in Discover Artificial Intelligence.
Different cancers demand different tissue features
Working with histological images involves a number of challenges. For example, adjacent regions of the same tumor can differ in cellular composition and tissue structure. In a digitized tissue section, pathologists assess the shape of cells and nuclei, the arrangement of tumor and immune cells, and the boundaries between a tumor and the surrounding tissue. Computer vision can help quantify these features and compare them across regions of a slide.
The features of interest depend on the clinical question. In lung cancer, algorithms help determine the tumor subtype based on tissue structure. In breast cancer, they assess tissue structure and the number of receptors—marker proteins used to select therapy—on specimens with special staining.
In colorectal cancer, one question is where immune cells are located relative to tumor foci. In prostate cancer, an important task is finding small malignant foci in a biopsy image.
Five tools compared against expert annotations
As part of the review, the authors compared five tools on publicly available tissue images spanning four cancer types. They compared the regions each tool identified with expert annotations of cell nuclei. All tools used standard settings without dataset-specific tuning.
Cellpose and StarDist in Fiji use deep learning. CellProfiler and Icy used pixel intensity thresholds in this comparison and could mark additional tissue regions as nuclei. Ilastik, trained on a small number of annotated examples, in some cases merged adjacent nuclei.
Cellpose had the highest mean Dice score in four of the six comparisons, including the pooled MoNuSAC result. The Dice score measures how closely a tool's segmentation matches the expert annotations (1 means complete agreement). On colon cancer images, Cellpose and StarDist both scored 0.74 when rounded to two decimal places. On images of triple-negative breast cancer with a low cell density, the mean Dice score was 0.84 for StarDist and 0.82 for Cellpose. StarDist sometimes missed nuclei in densely packed tissue.
Clinical use requires scrutiny across laboratories
The authors suggest examining which regions of a slide contributed to the model's output. In prostate tissue, for example, a pathologist can check whether the model highlighted structures used to assess tumor grade. Assessing clinical utility requires studies across multiple institutions and evaluation of how AI support affects pathologists' decisions.
"You need to start not with an algorithm, but with the question that the doctor will have to answer based on the histological specimen," says Svetlana Illarionova, the review's first author, assistant professor and head of research group at the Skoltech AI Center.
"In some cases, it's important not to miss a tiny focus; in others, it's important to see exactly how different cells are arranged in the section. And in any case, the doctor must see which areas of the image the model is based on. Its reliability should be tested using materials from different laboratories, where they are stained and scanned in different ways."
Skoltech co-authors include junior research scientist Anastasiia Studenikina, Ph.D. student Margarita Zapevalina, Professor Alexander Bernstein and Assistant Professor Maxim Sharaev, who co-leads the joint Skoltech–University of Sharjah lab BIMAI-Lab.
The other co-authors are Joaquim Carreras of the Department of Pathology at Tokai University School of Medicine and Rifat Hamoudi, a professor at the University of Sharjah and the university's co-lead of BIMAI-Lab.
More information
Svetlana Illarionova et al., Review of computer vision algorithms in histopathology to decipher complex features in cancer, Discover Artificial Intelligence (2026). DOI: 10.1007/s44163-026-02592-3
Who's behind this story?
Sadie Harley
BSc Life Sciences & Ecology. Microbiology lab background with pharmaceutical news experience in oil, gas, and renewable industries. Full profile →
Robert Egan
Bachelor's in mathematical biology, Master's in creative writing. Well-traveled with unique perspectives on science and language. Full profile →
Citation: Researchers compare AI tools for analyzing tumor tissue (2026, November 18) retrieved 18 November 2026 from https://medicalxpress.com/news/2026-11-ai-tools-tumor-tissue.html
This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no part may be reproduced without the written permission. The content is provided for information purposes only.
来源:Medical Xpress · medicalxpress.com