PaperlayerView exact revision on arXiv

Source-linked AI summary

AI-assisted mitotic counting improves reproducibility and efficiency across multiple tumour types

Simon Graham, Mostafa Jahanifar, Quoc Dang Vu, Vygante Maskoliunaite, Donatas Petroska, Ruta Barbora Valkiuniene, Ayat Gamal Lashen, Jen Hong Ong, Amede Ogechi Nnorom, Sinclair Couper, Natasha Kardasz, Reshma Agrawal, Brinder Singh Chohan, Jose Luis Solorzano Rendon, Shonali Natu, Arvydas Laurinavicius, Nasir Rajpoot, David Snead

arXiv:2610.01813v1cs.AI

TL;DR

Mitotic counting is clinically important but can vary between pathologists, and evidence across human tumour types is limited. MitPro supports hotspot selection and mitotic-figure review while leaving region selection and final counts to pathologists. In the study, agreement and efficiency improved, with score changes more often upward than downward.

  • Problem

    Mitotic counting varies between pathologists, and clinical evidence for AI-assisted counting across routinely assessed human tumour types remains limited.

    Source: Introduction · p. 2

  • Method

    MitPro displays detected mitotic figures and proposed hotspots, while pathologists retain control over region selection, corrections and the final count.

    Source: AI-assisted mitotic counting · p. 6, Discussion · p. 15

  • Results

    Agreement and assessment efficiency improved across tumour groups; scores were unchanged in 64.9% of assessments, higher in 29.1% and lower in 5.9%.

    Source: Impact on mitotic count and score · p. 13, Discussion · p. 15

  • Takeaways & Limitations

    The findings support MitPro as a clinically useful decision-support tool for more consistent and efficient mitotic assessment.

    Source: Discussion · p. 16

    The clinical study assessed reproducibility and efficiency, not accuracy, and established no independent ground truth for its slides.

    Source: Discussion · p. 15

Abstract

from arXiv · show

Mitotic counting is an important component of tumour grading, diagnosis and prognostic assessment across several tumour types, but manual assessment is time-consuming and subject to inter-pathologist variability. To help address these challenges, we developed MitPro, an AI tool designed to improve consistency and efficiency by directing pathologists towards regions with the highest predicted mitotic activity and highlighting mitotic figures for review, while retaining pathologist control over region selection and the final count. We evaluated its effect on the reproducibility and efficiency of mitotic counting in a retrospective, non-interventional, paired reader study comprising 385 whole-slide images from 3 centres in 3 countries and 7 tumour types using 3 different scanners. 13 pathologists participated, with each slide assessed independently by 3 pathologists without AI assistance and again with AI assistance after a minimum 2 week washout period. Across all slides, AI-assisted counting increased the intraclass correlation coefficient from 0.589 to 0.949. Mean pathologist-level median assessment time decreased from 286.4 to 127.8 seconds, corresponding to an average saving of 151.8 seconds per assessment. Improvements in agreement and efficiency were also observed in supporting analyses using HALO AP and Sectra image management systems and in 2 additional tumour types outside the main study population. AI-assisted assessment was associated with a subtle shift towards higher mitotic counts and scores, consistent with identification of more active mitotic hotspots and fewer missed mitotic figures. The frequency of score change between unassisted and AI-assisted assessment was comparable with inter-pathologist variation during routine counting. These findings support the use of MitPro as an assistive tool for more consistent and efficient mitotic assessment in routine practice.

Introduction

Mitotic counting is clinically useful but varies between pathologists, in part because hotspots and individual mitotic figures can be difficult to identify consistently. MitPro combines whole-slide detection with pathologist-guided hotspot selection and figure review, addressing a need for broader clinical validation across human tumour types.

Introduction

  • Mitotic activity varies within heterogeneous tumours, so pathologists assess hotspots; reporting mitoses per mm² is recommended to improve comparability across centres.The recommendation addresses differences in microscope high-power-field area, particularly as routine assessment becomes more digital.Source: Introduction · p. 2
  • Hotspot selection and recognition of individual mitotic figures are likely sources of inter-pathologist variation, as mitoses can be sparse, overlooked or confused with mimics.Examples of mimics include apoptotic bodies, pyknotic nuclei and crushed or degenerate cells.Source: Introduction · p. 2
  • Whole-slide AI detection can direct pathologists to regions with the most detected mitotic figures and display detections for review, supporting hotspot selection and figure recognition.Digital pathology also permits fixed-area regions to be defined on screen, supporting counting over a known area.Source: Introduction · p. 2
  • Clinical evidence for AI-assisted counting across the broader range of routinely assessed human tumour types remains limited, motivating multicentre evaluation across seven types.Prior work included selected high-power fields or specific tumour settings, alongside reports of improved agreement, accuracy or efficiency.Source: Introduction · p. 2

Materials and methods

The study used a paired, two-stage reader design to assess MitPro-assisted counting against unassisted practice on a multicentre, multi-tumour dataset. MitPro presented detections and proposed hotspots for review, while pathologists retained control of regions, corrections and final counts.

Study design

  • A retrospective, non-interventional, multicentre, multi-reader, two-stage paired study evaluated agreement and efficiency under conditions intended to represent routine diagnostic workflows.The same three randomly selected pathologists assessed each slide in both stages, separated by at least two weeks to minimise recall bias.Source: Study design · pp. 3–4

Dataset

  • The primary analysis included 385 unique patients from three centres in three countries, with slides scanned on three platforms at 0.23–0.50 µm/pixel.Cases covered seven tumour types; breast carcinoma included resections and biopsies, while other types included resections only.Source: Dataset · p. 4

Mitotic figure detection

  • MitPro's detection model was evaluated on an independent set of 463 regions and achieved an F1 score of 0.803, with precision 0.728 and recall 0.894.The independent validation set comprised 1,827 consensus mitotic figures across 215 whole-slide images from three centres.Source: Mitotic figure detection · pp. 5–6

Pathologist onboarding

  • All 13 primary-analysis pathologists completed onboarding and practice with the protocol and MitPro workflow before reporting study counts.Four additional pathologists joined the supporting analyses, bringing the full-study total to 17.Source: Pathologist onboarding · p. 6

Unassisted mitotic counting

  • In Stage 1, pathologists counted manually according to their usual clinical practice, without AI overlays or hotspot suggestions.Counts used tumour-specific reporting areas: 2 mm² for breast carcinoma, thyroid tumours, GI NETs and sarcoma, and 1 mm² for melanoma, meningioma and leiomyoma.Source: Unassisted mitotic counting · p. 6

AI-assisted mitotic counting

  • In Stage 2, MitPro displayed detected figures and predicted hotspots matched to each tumour type's reporting area, while pathologists could select custom regions, correct detections and confirm the count.The proposed regions and counts supported the assessment but did not determine the final result.Source: AI-assisted mitotic counting · p. 6

Statistical analysis plan

The statistical plan centred on individual-reader agreement in mitotic counts, with paired analyses of score agreement and assessment time. It also examined whether AI assistance was associated with shifts in counts or scores relative to routine reader variability.

Statistical analysis plan

  • The primary endpoint was the one-way random-effects ICC(1,1), estimating agreement between individual pathologists; stage differences and confidence intervals used paired cluster bootstrap resampling.Slides were resampled with the same indices for both stages to preserve pairing; simulations informed a target of approximately 40 cases per tumour group.Source: Statistical analysis plan · pp. 6–7
  • Agreement in tumour-specific mitotic scores was assessed with Fleiss’ kappa, with paired bootstrap confidence intervals for stage estimates and their difference.Mitotic scores use tumour-specific thresholds and represent one component of the wider pathological assessment, not necessarily determining grade or diagnosis.Source: Statistical analysis plan · p. 7
  • Assessment time was compared using paired observations from the same pathologist and slide, while count and score shifts were compared between matched stages and contextualised against routine reader variation.Stage-level time summaries averaged each pathologist's median; time saved was calculated from paired assessment-level differences.Source: Statistical analysis plan · p. 7

Results

AI assistance increased inter-pathologist agreement and reduced assessment time. It was associated with more use of proposed hotspots and a modest shift towards higher mitotic counts and scores, alongside greater scoring consistency.

Interobserver agreement in mitotic count

Interobserver agreement on mitotic scores

  • Fleiss’ kappa increased across tumour groups; improvements were significant for breast resection, breast biopsy, thyroid, gastrointestinal neuroendocrine tumours and sarcoma, but not meningioma or leiomyoma.Source: Interobserver agreement on mitotic… · p. 10

Mitotic count assessment time

  • Mean pathologist-level median assessment time fell from 286.4 to 127.8 seconds, saving 151.8 seconds per assessment, with reductions observed across every tumour group.Source: Mitotic count assessment time · p. 11

AI-assisted hotspots

  • Pathologists selected a MitPro-proposed hotspot in 87.2% of assessments and a manual region in 12.8%.Custom hotspots were used by at least some pathologists across all tumour types.Source: AI-assisted hotspots · p. 13

Impact on mitotic count and score

  • Across combined dataset assessments, scores were unchanged in 64.9%, higher in 29.1%, and lower in 5.9%; stage-to-stage score changes (64.9% unchanged) were comparable to Stage 1 between-pathologist variation (65.7% same-score pairwise comparisons).Source: Impact on mitotic count and score · pp. 13–14

Discussion

The discussion reports improved agreement and reduced assessment time with MitPro across tumour groups, linking these gains to support for hotspot selection and mitotic-figure recognition while retaining pathologist control.

Discussion

  • Agreement between pathologists and assessment time improved across all eight tumour groups, with gains in categorical mitotic-score agreement and consistency across centres, demographic groups and scanners.Pairwise correlations showed the same pattern.Source: Discussion · p. 15
  • MitPro supports hotspot selection and recognition of individual mitotic figures, addressing both potential sources of counting variability while leaving the final count to pathologists.Source: Discussion · p. 15
  • Pathologists selected a proposed hotspot in most AI-assisted assessments, but manual regions remained necessary in a minority, supporting the option to override unsuitable suggestions.Source: Discussion · p. 15
  • AI assistance was associated with higher counts in some tumour groups; score changes were more often upward, were not significant, and occurred at a frequency comparable with routine inter-pathologist differences.The upward shift was consistent with improved detection of active regions and mitotic figures, but does not by itself establish accuracy.Source: Discussion · p. 15
  • The study assessed reproducibility and efficiency without independent ground truth, and its three-reader design, group-level power and incomplete technical coverage limit definitive subtype-specific conclusions.Findings for less common subtypes are supportive rather than definitive.Source: Discussion · pp. 15–16
  • Pathologists remain responsible for selecting regions, reviewing detections and confirming counts; laboratories should recognize automation-bias risk and use quality-control practices to monitor it.Suggested monitoring includes test slides with challenging mitotic mimics and known counts.Source: Discussion · pp. 15–16
  • Agreement, correlation and assessment-time improvements were reproduced in HALO AP and Sectra despite differences in interface and verification workflow.Source: Discussion · p. 16
  • Similar benefits across 76 glioblastoma and non-small cell lung cancer slides suggest MitPro’s effects extend beyond the primary tumour types.The primary dataset covered three centres in three countries, seven tumour types and three scanner models, with the same readers assessing each slide in paired stages.Source: Discussion · p. 16

Declaration of competing interests

The competing-interests declaration identifies Histofy Ltd. roles and shareholdings for three authors.

Declaration of competing interests

Ethical Approval

The study received ethical approval from the London - Fulham Research Ethics Committee in February 2025.

Ethical Approval

Supplementary Material S1 Primary Analysis S1.1 Dataset The eight study groups were represented by similar sample numbers, with breast biopsies and resections analysed separately. The distribution is shown in Figure S1.

Across the primary analyses, AI assistance was associated with higher inter-pathologist agreement and correlation, shorter assessment times, and frequent use of proposed hotspots. Mitotic counts and scores shifted slightly upward, while most scores remained unchanged.

Supplementary Material S1 Primary Analysis S1.1 Dataset The eight study groups were represented by similar sample numbers, with breast biopsies and resections analysed separately. The distribution is shown in Figure S1.

S1.2 Interobserver agreement in mitotic count

S1.3 Bonferroni-adjusted analysis on interobserver agreement

  • Five tumour groups met the Bonferroni-adjusted superiority criterion; the other three had positive ICC differences but did not meet that stricter threshold.For breast biopsy, gastrointestinal neuroendocrine tumours and meningioma, adjusted confidence intervals included zero; all three nevertheless showed improvement at the conventional 95% confidence level.Source: S1.3 Bonferroni-adjusted analysis on… · p. 23

S1.4 Interobserver agreement on mitotic scores

S1.5 Pairwise pathologist correlation

  • Across the combined dataset, pooled pairwise correlation rose from 0.764 during routine counting to 0.975 with AI assistance, with increases across tumour groups and centres.The between-stage confidence interval remained above zero in every tumour group and centre.Source: S1.5 Pairwise pathologist correlation · p. 24, Table S3 · p. 25

S1.6 Mitotic count assessment time

  • Across 385 slides, mean pathologist-level median assessment time fell from 286.4 to 127.8 seconds, saving an average of 151.8 seconds per assessment.The largest tumour-group savings were 214.6 seconds for sarcoma and 208.8 seconds for meningioma.Source: S1.6 Mitotic count assessment time · p. 27, Table S5 · p. 27

S1.7 AI-assisted hotspots

  • Pathologists selected a MitPro-proposed hotspot in 87.2% of assessments, with use ranging from 81.2% for thyroid tumours to 92.2% for meningioma.A manually defined alternative hotspot was used in the remaining 12.8% of assessments.Source: S1.7 AI-assisted hotspots · p. 29, Table S7 · p. 29

S1.8 Impact on mitotic count and score

  • Mitotic scores mostly remained unchanged between stages; among changes, increases were more common than decreases, and Stage 2 mean counts stayed within one Stage 1 standard deviation in all eight groups.Counts shifted slightly upward overall, with larger shifts generally in higher-activity groups; the study did not establish the cause of individual score changes.Source: S1.8 Impact on mitotic count and score · p. 30, Table S9 · p. 30, Figure S9 · p. 31

S2.1 Performance across image management systems

Supporting analyses found that agreement and efficiency gains were reproduced within both HALO AP and Sectra workflows. Additional analyses in glioblastoma and NSCLC also showed improved agreement and shorter assessment times, although those tumour types were outside MitPro’s intended use.

S2.1.1 Overview

  • Across 95 slides assessed in both systems, within-system comparisons—not direct comparisons between systems—were used to evaluate Stage 1 and Stage 2.All tumour groups from the main study were represented, and results were pooled across groups.Source: S2.1.1 Overview · p. 34

S2.1.2 Interobserver consistency and correlation

  • ICC increased from 0.672 to 0.965 in HALO AP and from 0.608 to 0.962 in Sectra; pooled pairwise correlations also rose in both systems.The between-stage ICC confidence intervals were above zero in each system; reader groups differed, so the results do not establish a performance difference between systems.Source: S2.1.2 Interobserver consistency and… · p. 34, Figure S12 · p. 35

S2.1.3 Assessment time

S2.2.1 Overview

  • The outside-intended-use analysis included 76 slides: 38 glioblastoma cases and 38 NSCLC cases, assessed using raw counts because applicable clinical score cut-offs were unavailable.These results provide supporting evidence of broader generalisability, not support for extending MitPro’s intended use.Source: S2.2.1 Overview · p. 36

S2.2.2 Interobserver consistency and correlation

S2.2.3 Assessment time

S3 Mitotic figure detection performance

Analytical validation compared MitPro detections with consensus pathologist annotations across tumour groups, subtypes and technical conditions. Performance was broadly maintained across represented groups and several image perturbations, with lower estimates in some small subgroups and under stronger compression or a stain target.

S3.1 Dataset

  • The 215-slide dataset covered seven tumour types, and 1,827 consensus annotations were constructed from independent pathologist markings in predefined regions.The validation also included regions without mitotic figures, non-tumour tissue and common artefacts.Source: S3.1 Dataset · pp. 38–39

S3.2 Performance across tumour types

  • Across all tumour types and centres combined, MitPro was evaluated against consensus pathologist annotations across the complete analytical validation dataset, achieving precision 0.728, recall 0.894, and F1 score 0.803.Source: S3.2 Performance across tumour types · p. 40

S3.3 Performance across subtypes

  • Subtype F1 scores varied: breast lobular carcinoma scored 0.884 versus 0.722 for infiltrating duct carcinoma, while rare sarcoma categories scored 0.533 and 0.516.The lower rare-sarcoma estimates and some other subtype results were based on small numbers of slides or detections.Source: S3.3 Performance across subtypes · p. 42, Figure S17 · p. 43

S3.4 Performance across centre, sex, age and scanner

S3.5 Generalisation to additional cancer types

  • On the additional glioblastoma and NSCLC dataset, MitPro achieved an F1 score of 0.794, with scores of 0.784 and 0.809 for glioblastoma and NSCLC, respectively.Some age-group variation was observed, particularly for the smaller under-40 group, which had wider confidence intervals.Source: S3.5 Generalisation to additional… · p. 45, Table S24 · p. 46

S3.6 Robustness analysis

  • F1 was broadly maintained under blur and noise and at moderate JPEG compression, while the strongest compression and one stain target produced the largest changes.F1 was 0.721 at JPEG quality 25 versus 0.746 for original images, and ranged from 0.679 to 0.724 across stain targets versus 0.746 original.Source: S3.6 Robustness analysis · pp. 47–48, Figure S22 · p. 49

Related papers

Loading 2610.01813v1…