The Pulse
MitPro Raises Pathologist Agreement Across 385 Tumor Slides
A study of 385 whole-slide images found that pathologists’ agreement on mitotic counts increased when they used MitPro, an AI tool that suggests hotspot regions and candidate mitotic figures. Assessment time also fell, though the study focu
AI.info Team · Written with AI ·
On October 1, researchers posted results from a three-country study in which AI assistance sharply increased agreement among pathologists counting dividing cells in tumor samples. Across 385 whole-slide images, the agreement measure rose from 0.589 during unaided review to 0.949 with MitPro, a tool that suggests active regions and flags candidate mitotic figures for human review.
The study also found that pathologists spent less time on each assessment when using the software. The results cover seven tumor types and three scanners, but they do not show that MitPro can diagnose cancer or determine a patient’s prognosis: pathologists kept responsibility for the final count, and the study measured agreement and time rather than patient outcomes.
Three readers assessed each of 385 slides
The study included 385 whole-slide images from 385 unique patients at three clinical centers in Lithuania, the United Kingdom and Spain. The images covered breast carcinoma, melanoma, thyroid tumors, meningioma, gastrointestinal neuroendocrine tumors, sarcoma and leiomyoma. Researchers used three scanner models, and the slides were prepared under each laboratory’s staining protocols.
Thirteen pathologists participated in the primary analysis, but three independently assessed each slide in each round. In the first round, readers counted without AI; in the second, they used MitPro after at least a two-week interval. The same three readers assessed a given slide in both rounds, allowing a paired comparison of their work.
Mitotic figures are cells caught in the process of division. Counting them in a tumor sample helps pathologists assess proliferative activity and contributes to grading in several cancers. The task can vary between readers because they must first locate a tumor’s most active region, then distinguish mitotic figures from lookalikes such as damaged or dying cells.
Agreement rose as assessment time fell
The researchers report that the overall intraclass correlation coefficient—a measure of agreement among readers—increased from 0.589 without AI to 0.949 with assistance. The improvement was reported across all seven tumor groups. Mean pathologist-level median assessment time fell from 286.4 seconds to 127.8 seconds; the authors calculate an average saving of 151.8 seconds per assessment.
MitPro scans the whole slide to propose hotspot regions and display candidate mitotic figures. Pathologists can choose another region, correct detections and confirm the count themselves. The model was trained on 28,222 annotated mitotic figures, while a separate analytical validation used 463 regions from 215 slides; the authors report an F1 score of 0.803 for that detection test.
“These findings support the use of MitPro as an assistive tool for more consistent and efficient mitotic assessment in routine practice.”
Simon Graham and co-authors, paper authors
Higher counts need human review
AI-assisted assessments showed a limited shift toward higher mitotic counts and scores, which the authors link to finding more active hotspots and fewer missed figures. The frequency of score changes was comparable to differences between pathologists during routine counting. The paper says the study was designed to assess reproducibility and efficiency, not accuracy, and did not establish an independent ground truth for the clinical slides.
The authors also note limits to the evidence. Three readers per slide may not capture the full range of differences among pathologists, and the study was powered around broad tumor groups rather than individual histological subtypes. Its sample cannot represent every laboratory process, specimen type or technical setup.
Histofy ties to the study warrant attention
MitPro is developed by Histofy Ltd. The paper says the work received support from Histofy and Innovate UK; it also discloses that two authors are Histofy directors and shareholders and another author holds shares in the company.
The paper describes MitPro as decision support, not an autonomous counter. Pathologists remain responsible for selecting the counting region, reviewing detections, correcting errors and confirming the final count. The authors also note that familiarity with decision-support tools may lead readers to rely more on suggestions and become less likely to challenge incorrect ones. For this study, the practical finding is narrower: in the tested slides and reader workflow, pathologists agreed more often and finished their counts faster when MitPro assisted them.