AI made three pathologists more consistent on bladder-cancer grade—but prospective evidence is missing
Across 465 retrospective cases, model-assisted grading reduced disagreement and raised average discrimination. The external accuracy gain did not reach statistical significance, and no treatment or patient outcome was tested.
The short version
- What happened
- The model was trained on 91,183 image patches with 318,893 annotations; three pathologists then graded 149 internal and 316 external cases with and without assistance.
- Who it affects
- Patients may benefit if a reproducible mitotic count reduces avoidable grading disagreement, but the study does not show that an AI-assisted grade improves treatment selection or outcomes.
- What remains unproven
- Only three pathologists participated
Keep in mind
Only three pathologists participated
Retrospective crossover design may not reproduce routine workflow
External AUC improvement was not statistically significant
No treatment, safety, recurrence or survival endpoint
Source analysis · 1 linked source. Primary analysed source date: 22 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
The strongest result was less disagreement between readers
Researchers trained a whole-slide mitosis detector on 91,183 image patches containing 318,893 annotations. Three independent pathologists then completed a crossover exercise on 465 urothelial-carcinoma cases: 149 from an internal cohort and 316 from external hospitals, first grading without the model and then with its mitotic count.
The model's count alone distinguished high-grade disease with an area under the receiver-operating-characteristic curve of 0.91 (95% confidence interval 0.89–0.94). Assistance raised mean reader AUC from 0.90 to 0.98 internally. Externally it rose from 0.84 to 0.95, but the p value of 0.11 means the study could not rule out chance for that accuracy difference. Agreement improved in both cohorts: Fleiss's kappa rose from 0.72 to 0.93 internally and from 0.62 to 0.85 externally.[1]
A useful second opinion is not yet a clinical outcome
For a patient, tumour grade can influence treatment intensity and follow-up. More consistent measurement could therefore matter, but the study retrospectively reused digitised cases in a controlled reader exercise. It did not test whether pathologists work faster, whether disagreements are resolved safely or whether assisted grading improves treatment choices or outcomes.
The authors call for prospective validation. A convincing next study would pre-register thresholds, enrol consecutive cases from varied laboratories and scanners, measure errors and turnaround time, and follow how model-assisted grades alter multidisciplinary decisions. The work received Korean government health-research funding; the authors declared no competing interests.[1]
The distinction that matters
What the reader study actually compared
The same three-reader design was applied to an internal and an external cohort.
Internal: 149 cases
Grade without and with AI
AUC 0.90 → 0.98; kappa 0.72 → 0.93
External: 316 cases
Grade without and with AI
AUC 0.84 → 0.95; kappa 0.62 → 0.85
Clinical deployment
Not tested
No treatment or patient outcomes