Does AI help pathology residents read prostate biopsies more accurately?
In a peer-reviewed study, 15 residents improved when AI annotations were shown after an unaided read of the same 186 slides. The fixed order means the result cannot separate AI assistance from a second look, and the developer employed four authors.
The short version
- What happened
- Fifteen PGY-1 to PGY-3 pathology residents each reviewed 186 whole-slide images: 104 adenocarcinomas, 75 benign slides and seven atypical small acinar proliferations.
- Who it affects
- Residents may receive useful visual cues on difficult slides, but training should preserve an unaided judgment and supervised discussion rather than encourage automatic acceptance.
- What remains unproven
- Fifteen residents and 186 selected slides, including only seven atypical cases
Keep in mind
Fifteen residents and 186 selected slides, including only seven atypical cases
Fixed unaided-then-assisted order confounds AI assistance with a second look, memory and learning
No practising-pathologist comparison, retention test, workload measure or patient outcome
Four authors worked for the platform and model developer
Single task and system; generalisability across centres, scanners and populations is untested
Source analysis · 1 linked source. Primary analysed source date: 4 November 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 9 min The evidence, methods, implications and limitations
The direct answer: assisted readings were more accurate, but the design cannot isolate why
In this study, pathology residents produced more correct prostate-biopsy interpretations after an AI system displayed tumour regions and Gleason-pattern annotations. The improvement was largest among residents with less previous genitourinary pathology experience. That is a meaningful result for decision support and training, because prostate biopsies require readers to distinguish benign tissue, atypical small acinar proliferations and adenocarcinoma while grading cancer patterns that affect care. It is not proof that the AI alone caused the full improvement.
All 15 residents reviewed the same 186 whole-slide images in a fixed sequence: first unaided, then with AI assistance. The second read therefore had two advantages at once—the annotations and another opportunity to inspect a familiar slide. Memory, reconsideration and learning can all improve a second answer even without AI. A counterbalanced trial, in which some slides are assisted first and others unaided first, would separate those effects more convincingly. The authors acknowledge this directly and describe educational value as a hypothesis rather than a measured outcome.[1]
Who read what, and what the system showed
The prospective within-subject study enrolled 15 pathology residents from postgraduate years one to three. The dataset contained 186 prostate core-biopsy whole-slide images: 104 adenocarcinomas, 75 benign slides and seven atypical small acinar proliferations. Each resident generated both an unaided and an assisted interpretation, creating repeated observations across readers and slides. Residents were grouped by prior genitourinary pathology experience rather than randomly assigned to experience levels. The platform highlighted tumour regions and provided Gleason-pattern annotations intended to draw attention to malignant tissue and support grading.
The denominator is important. The study evaluates 15 trainees, not practising pathologists, and 186 selected images, not a prospective stream of patients with all the distractions, missing context and time pressure of routine service. Seven atypical slides are too few to support a confident subgroup claim about that difficult category. The paper evaluates slide-level diagnostic judgments rather than treatment decisions or patient outcomes. It also does not measure how long each review took, whether residents relied on an incorrect annotation or whether assistance changed the quality of their written explanation.[1]
The distinction that matters
What the study establishes—and what it leaves unresolved
The paired readings show better performance when AI was visible, but fixed ordering prevents a clean causal estimate of the AI's contribution.
Participants
15 PGY-1 to PGY-3 residents
Trainee evidence, not expert-practice evidence
Cases
186 whole-slide images
104 cancer, 75 benign, 7 atypical
Comparison
Unaided first, AI-assisted second
AI and second-look effects are mixed
Outcome
Accuracy, sensitivity and specificity
Improved assisted performance; no retention or patient outcome
The observed improvement was larger for less-experienced residents
Among residents classified as less experienced, observed accuracy increased from 62.3% unaided to 75.1% with the display, a gain of 12.9 percentage points with a reported 95% interval from 7.8 to 18.7. Among more-experienced residents, accuracy increased from 77.9% to 83.0%, a 5.1-point gain with an interval from 3.2 to 7.5. Bayesian hierarchical mixed-effects models estimated that AI assistance increased the odds of a correct diagnosis by 2.39, with a 95% credible interval from 1.94 to 2.96.
The model also estimated higher odds of cancer detection, at 2.55 with a credible interval from 1.83 to 3.62, and correct benign classification, at 2.76 with a credible interval from 2.04 to 3.74. The interaction suggested a smaller relative benefit for more-experienced residents on overall accuracy and specificity. It did not establish a clear experience difference for sensitivity: the interaction estimate was 0.82 and its credible interval, 0.51 to 1.37, crossed one. That distinction matters because the headline finding of a narrowed experience gap is stronger for some outcomes than for every component of performance.[1]
The error transitions show benefit and new mistakes at the same time
Across the paired readings, the AI display converted 307 of 786 unaided errors into correct diagnoses—39% of the initial errors. It also changed 75 of 1,913 correct unaided diagnoses into errors, or 3.9%. Those two numbers are more informative together than a single accuracy average. Assistance can rescue a missed cancer or mistaken benign call, but it can also persuade a reader away from a correct judgment. A clinical implementation therefore needs to measure harmful reversals as well as corrected mistakes and should show when the resident or pathologist disagrees with the system.
Exploratory analysis suggested larger benefit on slides that were more difficult during unaided review. The authors used an ability-weighted difficulty measure derived from unaided performance and included it as a covariate. That result is consistent with a ceiling effect: easy slides leave little room to improve, while difficult slides create more opportunity for useful cues. It is exploratory and partly defined from the same unaided performance being analysed. Independent difficulty ratings, pre-specified categories and external cases would strengthen the claim that assistance is especially useful on genuinely difficult material.[1]
Why better second answers do not yet prove better education
A decision-support tool can improve a task without teaching the user to perform better later. The study did not remove the display and retest residents on new slides after a delay, so it cannot show retention, transfer or independent skill development. If trainees learn to compare their own reasoning with a transparent annotation, the tool could accelerate pattern recognition. If they mainly follow highlighted regions, it could improve immediate scores while weakening attention to features the system misses. Both outcomes are plausible; neither is measured here.
A training study should distinguish feedback from answer substitution. Useful measures would include performance on unseen slides after assistance is withdrawn, calibration of confidence, ability to explain the diagnosis, time to decision and patterns of disagreement. It should also examine whether residents notice false-positive highlights and false-negative regions. A counterbalanced or randomised design could vary when assistance appears, while a longitudinal design could test whether any gain persists. Until then, the evidence supports assisted performance during the displayed review, not a claim that AI has made residents more capable pathologists.[1]
Developer involvement and the reference standard need visible scrutiny
Four authors were affiliated with NovinoAI, the developer of the FlexLIS digital-pathology platform and the associated models evaluated in the study. The remaining authors declared no competing interests, and the paper reports no external funding. Developer participation does not invalidate a result, but it increases the importance of preregistered analysis, independent replication, accessible data and testing by teams without a commercial stake. The article says deidentified tissue and data were used under independent institutional review-board approval, with consent from participating residents and a waiver for the deidentified patient material.
The published abstract and methods describe Bayesian hierarchical models fitted with Markov chain Monte Carlo to account for repeated observations and variation between readers and slides. That is appropriate for a clustered within-subject dataset, but statistical sophistication cannot repair the fixed reading order. The reference diagnoses, slide selection process, platform configuration and thresholds also influence apparent performance. A hospital considering the tool would need local validation across scanners, staining variation, specimen quality and patient groups, plus a governance plan for model changes after deployment.[1]
What this means for trainees, pathologists and patients
For residents, the safest interpretation is that a visual decision aid may help expose overlooked tumour regions and reduce some experience-related performance gaps during review. It should not replace supervision, independent reasoning or exposure to difficult negative cases. Training programmes can record the resident's unaided answer before revealing assistance, then discuss disagreements with a supervising pathologist. That preserves a learning signal and reduces the risk that the system's suggestion becomes the only visible answer.
For practising pathologists, the result is not direct evidence of improved routine service because experts were not the target population. For patients, no outcome was measured: the study does not show fewer missed cancers, better grading decisions, faster treatment or reduced harm. Before clinical use, evaluation should count false reassurance, unnecessary escalation, grade changes, turnaround time, downstream review and cases in which an incorrect annotation influences the final report. Human sign-off remains essential, but it must be supported by workload and interface design that make disagreement realistic rather than nominal.[1]
Evidence limits and what would change the assessment
This is a peer-reviewed prospective study with a clear denominator, paired readings and uncertainty estimates. Those are strengths. Its main limitations are 15 residents, a fixed unaided-then-assisted sequence, selected whole-slide images, only seven atypical cases, no practising-pathologist comparison, no time or workload outcome, no retention test and no patient outcome. Four authors' affiliation with the platform developer adds a commercial interest that independent replication should address. The findings apply to this system, task and study setting; they should not be generalised to every pathology model or cancer type.
The assessment would become stronger with a randomised counterbalanced multi-centre trial, blinded independent reference review, more diverse slides and readers, predefined subgroup analyses and transparent reporting of false guidance. A longitudinal education study should show whether residents retain and transfer skills after assistance is removed. A clinical study should measure report accuracy, turnaround, corrections and patient-relevant consequences under normal workload. Today's paper is consequential evidence about supervised decision support, but it reports no urgent change in care and no active hazard. It is therefore Research & Impact coverage, not Breaking news.[1]
Watch the background, then check the evidence
Illustrative photo · not a video frame
The Royal Society · 44 min 28 sec · Background explainer
Is this the AI future we want? | The Royal Society
YouTube loads only when you choose; playback starts when you press play.
A Royal Society discussion with Madhumita Murgia and biomedical engineer Alison Noble on AI’s potential in healthcare, science and culture, recorded at the 30 June 2026 event.
Evidence check: A discussion of opportunities and risks, including future possibilities. It is not evidence that every application described is available or clinically proven.
Published 2026-07-01. Source checked 14 October 2026. Check captions on the publisher’s video page.
Check the healthcare evidencePhoto: Daniel Sone / National Cancer Institute / Unsplash · Licence
Continue the story
Read next
Continue exploring prostate cancer: compare the finding and its evidence limits.
Can prostate-pathology AI generalise to Iraqi patients?
Three research models trained outside Iraq showed strong agreement with pathologists on 339 biopsy slides from 185 patients in Erbil. The result supports local validation, not autonomous diagnosis or immediate clinical deployment.
Read the next reportExplore the full source and evidence record
What this means for people
- Residents may receive useful visual cues on difficult slides, but training should preserve an unaided judgment and supervised discussion rather than encourage automatic acceptance.
- Pathologists need interfaces that make disagreement, correction and audit visible, especially when an AI cue changes a previously correct answer.
- Patients were not outcome participants in this study; reduced missed cancer, better grading and faster care remain unproven.
Global context
Digital pathology systems are spreading across health systems, but training, scanner infrastructure, reference standards and regulatory oversight differ by country. Results from US pathology residents on one platform cannot be assumed to transfer to every service or cancer programme.
What the evidence does not yet show
- Fifteen residents and 186 selected slides, including only seven atypical cases
- Fixed unaided-then-assisted order confounds AI assistance with a second look, memory and learning
- No practising-pathologist comparison, retention test, workload measure or patient outcome
- Four authors worked for the platform and model developer
- Single task and system; generalisability across centres, scanners and populations is untested
What to watch next
- Counterbalanced, randomised multi-centre studies
- Independent replication without developer involvement
- Retention and transfer after AI assistance is removed
- Performance on atypical, rare and technically difficult specimens
- Clinical workflow, correction and patient-outcome evidence
Living evidence record
Impact record IAI-1RYRW8V
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
4 November 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what BMC Medical Education published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 4 November 2026
Research question, topics and regions
Whether AI annotations improve pathology residents' diagnostic accuracy, cancer sensitivity and benign-slide specificity when interpreting prostate core-biopsy whole-slide images
Topics: Digital pathology · Medical education · Prostate cancer · Decision support · Human-AI collaboration
Regions: United States · North America · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.