Industries / Health / Healthcare / Medical labs
Pathology slides
Cutting tissue removed from patients into thin slices, staining it on glass slides, and reading it under a microscope or on a screen to diagnose diseases such as cancer.
7
Milestones achieved, from 1905 to 2025.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1905
A diagnosis during surgery
Louis B. Wilson, appointed chief pathologist at Mayo Clinic in 1905, froze a piece of tissue removed during an operation, stained it and mounted it on a slide, which showed whether the growth was cancer while the patient was still in surgery. Mayo Clinic's history says he did it to answer William J. Mayo's request for such a method, and that its diagnostic rooms are still placed close to the operating rooms.
Credit
Louis B. Wilson, Mayo Clinic
1943
Cells read from a smear
George Papanicolaou and Herbert Traut, a gynecological pathologist, published Diagnosis of Uterine Cancer by the Vaginal Smear in 1943. It showed that normal and abnormal cells scraped from the vagina and cervix could be correctly sorted under the microscope. The test cost little and became the standard screen for cervical cancer, and its wide use was followed by a large fall in the number of cases.
Credit
George N. Papanicolaou and Herbert F. Traut, Cornell University and New York Hospital
2007
One in five tests suspect
HER2 is a protein that some breast cancers overproduce, and its level decides who gets drugs aimed at it. The American Society of Clinical Oncology (ASCO) and the College of American Pathologists (CAP) found that about 20% of HER2 testing may be inaccurate. Their guideline set fixed cutoffs for reading the stain, for example "3+" for strong, even staining of the cell edge in more than 30% of tumor cells, and required each lab to show 95% agreement with another validated test.
LessonA stain read by eye gives the same answer in every lab only with fixed rules, validation in each lab and outside checks.
Credit
Antonio C. Wolff, M. Elizabeth Hammond and the ASCO/CAP expert panel
2017
Programs beat a timed panel
In the CAMELYON16 challenge, teams built programs to find breast cancer that had spread to lymph nodes, using scanned slides from two Dutch hospitals, and were scored on 129 new slides. The best program scored 0.994 on a scale where 0.5 is chance and 1 is perfect. Eleven pathologists reading the same glass slides in a two-hour session averaged 0.810. The top programs did about as well as one pathologist working without a time limit.
Credit
Babak Ehteshami Bejnordi, Jeroen van der Laak and the CAMELYON16 Consortium
2021
Models learned the hospital
The Cancer Genome Atlas (TCGA), a US archive of tumor slides and genetic data, is one of the largest public sources of scanned slides. Using slides from more than 3,000 of its patients with six cancer types, researchers showed that a program could tell which hospital had sent each slide, even after common color corrections, and that this made models for survival, mutations and tumor stage look more accurate than they were.
LessonTest sets must come from hospitals whose slides the model never saw in training.
Credit
Frederick M. Howard, Alexander T. Pearson and colleagues, University of Chicago
2021
Software flags prostate cancer
On 21 September 2021 the US Food and Drug Administration (FDA) granted a De Novo request, its route for new kinds of low- to moderate-risk devices, for Paige Prostate. The software marks the spot most suspicious for cancer on a scanned prostate needle biopsy for the pathologist to check. On 171 slides with cancer in the main study, the FDA estimated that pathologists using the software would find cancer they had otherwise missed or deferred in 7.3% of biopsies with cancer, and that specificity, the share of benign slides correctly called benign, rose by 1.1 points. The FDA added that the benefit per patient would likely be "substantially lower".
Credit
Paige.AI; US FDA
2025
Bigger models, small gains
Foundation models are AI models pretrained on millions of unlabeled slide images and then adapted to many tasks. A benchmark built from clinical slides at three medical centers compared the public ones. All of them detected disease well, scoring above 0.9 on the 0.5-to-1 scale, but all did poorly at predicting which lung cancers would respond to immunotherapy, the best averaging about 0.6. The authors found no strong evidence that larger models or more pretraining slides gave better results, while performance on a tissue rose with that tissue's share of the training slides, most clearly for lung.
LessonSize alone did not make a pathology model better.
Credit
Gabriele Campanella, Chad Vanderbilt and colleagues, Icahn School of Medicine at Mount Sinai and Memorial Sloan Kettering Cancer Center
Watch for
AI reads frozen sections
CRISP is a model trained on more than 100,000 frozen sections from eight medical centers and tested on more than 15,000 slides in nearly 100 tasks. Its makers report that in a forward-looking study of more than 2,000 patients it directly informed surgical decisions in 92.6% of cases and cut pathologists' diagnostic workload by 35%.
Achieved whenthe study passes peer review and a hospital outside the eight centers reports the model's accuracy on its own operations.
Claimed by
Zihan Zhao, Fengtao Zhou, Hao Chen, Muyan Cai and 19 coauthors (arXiv preprint), October 2025
General AI pans across slides
GIANT lets a general-purpose AI model pan and zoom across a whole scanned slide, choosing where to look next, as a pathologist does. On 128 questions written by pathologists, GPT-5 driving GIANT, with five runs combined, answered 62.5% correctly, against 43.8% for TITAN and 37.5% for SlideChat, two models built for pathology. A pathologist who read 28 of its reasoning records found that it often fixed on a first guess and reinterpreted later findings to fit it. A community rebuild on GitHub, gigapixel-goblin, reran the five public task sets with GPT-5.2 and reports mixed results: 60.1% on the pathologist-written questions, higher than the paper on one task and lower on two others.
Achieved whena documented evaluation outside the authors reproduces the scores, ideally on slides not in the original set.
Claimed by
Thomas A. Buckley, Arjun K. Manrai and colleagues, Massachusetts General Hospital and Harvard Medical School (arXiv preprint), November 2025, revised June 2026
Open
01
A HER2-low answer key
A public answer key for HER2 stains at the low end. At least 300 scanned breast cancer HER2 slides from a public dataset, each scored 0, 1+, 2+ or 3+ by at least five breast pathologists working alone under the current ASCO/CAP rules, with every score published, not only the majority. Then score open-source AI scorers and general AI models against it, with the line between 0 and 1+ reported on its own.
The briefCredit · unclaimed02
Slide agents on fresh slides
200 new questions written by pathologists outside the GIANT group on CPTAC slides, newer public slides that are less likely to be in models' training data, run through GIANT alongside the public MultiPathQA questions. Use at least four driver models, including one with downloadable weights, and publish accuracy, cost per slide and the failure types a pathologist finds in 50 reasoning records.
The briefCredit · unclaimed
Open 01 · the brief
A HER2-low answer key
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public answer key for HER2 stains at the low end. At least 300 scanned breast cancer HER2 slides from a public dataset, each scored 0, 1+, 2+ or 3+ by at least five breast pathologists working alone under the current ASCO/CAP rules, with every score published, not only the majority. Then score open-source AI scorers and general AI models against it, with the line between 0 and 1+ reported on its own.
Why it is still open
"HER2-low" (1+, or 2+ without extra copies of the gene) now decides who can get newer HER2 drugs, and this is where pathologists disagree most. In a 2023 study, 18 breast pathologists from 15 institutions scored 170 biopsies; they all agreed on fewer than 1% of 1+ cases and on 25% of 0 cases, and the authors called the scoring "unacceptably discordant" for HER2-low treatment. Public sets such as HER2-IHC-40x (107 slides, 2025) give one score per region, from expert annotators, not each pathologist's separate score. A search found expert reference sets built for training and national quality programs, but none released for public use.
Built with what exists
the ACROBAT data set, 4,212 public slide images from 1,153 breast cancer patients, with H&E and immunohistochemistry stains for ER, PGR, HER2 and KI67 from the same tumors; QuPath, an open-source slide viewer, for scoring; pathology foundation models whose weights can be downloaded, such as UNI and Virchow2, to train a small scorer from a few hundred labeled slides; general models such as Claude, Gemini, GPT and Qwen3-VL for comparison. The advance is foundation models, which let a small team build a slide scorer without millions of labels. No patient contact; published images only.
First step
Download the ACROBAT HER2 slides, pick 50 that span 0 to 3+, and have two pathologists score them in QuPath to time the work per slide.
Done looks like
300 or more slides scored by five or more pathologists, with agreement statistics (Fleiss' kappa) and every score published under an open license with a DOI. At least three AI methods scored on 0 versus not-0, both against the majority and against each pathologist.
Open 02 · the brief
Slide agents on fresh slides
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
200 new questions written by pathologists outside the GIANT group on CPTAC slides, newer public slides that are less likely to be in models' training data, run through GIANT alongside the public MultiPathQA questions. Use at least four driver models, including one with downloadable weights, and publish accuracy, cost per slide and the failure types a pathologist finds in 50 reasoning records.
Why it is still open
The published results come from the group that built GIANT. Its question set has been public on Hugging Face since November 2025. A community rebuild, gigapixel-goblin, has rerun those questions with one model and posted mixed scores, but every question so far sits on TCGA, GTEx and PANDA slides, collections widely used to train pathology models. Searches found no questions written on slides outside them.
Built with what exists
MultiPathQA from Hugging Face; OpenSlide, an open-source library for reading slide files; CPTAC whole-slide images from The Cancer Imaging Archive, licensed CC BY 4.0; Claude, Gemini and GPT through their APIs; Qwen3-VL-235B, which GIANT's authors also tested, as the downloadable model. The advance is AI models that can look at many image crops in one session and decide where to look next.
First step
Download MultiPathQA, write a pan-and-zoom loop with OpenSlide, and run the first 20 pathologist-written questions with one model.
Done looks like
All 934 MultiPathQA questions and 200 new CPTAC questions run with four or more models, five runs each; the scores either fall within the paper's error bars or the gap is explained; a pathologist's labeled review of 50 records; code, prompts and costs published.