The experimental AI system was designed to turn prostate-biopsy slides into German pathology reports, and it scored much higher on malignancy detection than on detailed cancer grading. In the Munich test cohort, malignancy accuracy was 97.6% and its F1-score was 96.2%; grading accuracy was 87.2% and grading F1-score was 65.2%. Across three outside cohorts, malignancy F1-scores ranged from 91.9% to 95.4%, while grading F1-scores ranged from 41.3% to 50.1%. The work is an arXiv version 1 preprint dated 24 August 2026.
How the system was assembled
The framework operates at slide level: it takes digitized whole-slide images of prostate biopsies and is intended to produce reports in German. The researchers trained a domain-specific tokenizer and the vision-language model from scratch, using no pre-trained language components.
To assemble training material, an automated local language-model pipeline split composite laboratory-information-system reports into core-specific pairs linking each whole-slide image to text. It produced 17,344 pairs from 2,402 historical cases without manual annotation.
A manual audit of 200 randomly sampled gradings found no discrepancies in the disaggregated text or structured labels. Because this was a sample, it did not amount to a review of every extracted pair.
A clear split between detection and grading
For the internal evaluation, the researchers kept cases in separate training and test partitions rather than splitting individual biopsies across the two. Training contained 14,976 biopsies from 2,200 cases, while the test partition contained 2,368 biopsies from 202 cases. The split used case-level stratification.
On the Munich test set, the model’s malignancy specificity was 98.2%, sensitivity was 96.4%, and precision was 95.9%, alongside the 97.6% accuracy and 96.2% F1-score.
Detailed grading was less consistent. Accuracy was 87.2%, while macro precision was 65.8%, macro recall was 65.7% and macro F1-score was 65.2%. The paper notes that rare high-grade categories and merged Gleason groups affected the grading result.
Tumor-infiltration estimates were less precise
The model also estimated tumor infiltration, reporting both an extent in millimeters and a percentage. On average, the extent estimates differed from reference values by 2.9 millimeters across 211 samples; percentage estimates differed by 15.2% across 82 samples. The two estimates had Pearson correlations of 0.637 and 0.78 with the references, but the authors treated them as approximate rather than precise clinical measurements.
The picture outside Munich
Testing beyond Munich used three cohorts that contributed neither training data nor hyperparameter selection: 424 biopsy cores from 50 Heidelberg cases, 5,456 biopsies from Karolinska and 5,160 from Radboud.
With the described augmentation, malignancy F1-scores were 91.9% in Heidelberg, 95.0% in Karolinska and 95.4% in Radboud. The corresponding grading F1-scores were much lower: 41.3%, 50.1% and 46.2%.
Training used lighter and darker versions of H&E-stained image tiles, together with augmentation of image embeddings in latent space. In a comparison with an otherwise identical model without augmentation, grading F1-score was higher by 8.7 percentage points on Karolinska, 1.6 on Radboud, 1.5 on Heidelberg and 4.2 on Munich. The comparison was not randomized, so it does not establish that augmentation caused the observed differences.
Benchmarks offer a mixed picture
Researchers also compared the report-generation system with an ABMIL classifier trained on exactly the same partitions. The malignancy F1-scores differed by 0.2%, while the report-generation model’s grading F1-score was 1.7% higher. The comparison was nonrandomized, so it cannot show that the report-generating format caused the difference.
Against an FDA-cleared comparison tool, the report-generation model’s malignancy F1-score was 1.1 percentage points higher. The tool’s malignancy F1-score was 95.1% and its grading F1-score was 68.8% on the evaluation dataset. Because the tool’s training dataset was not known, these benchmark figures do not establish that one system is superior.
PRISM, a model using pre-trained language components, recorded a malignancy F1-score of 94.5% and a grading F1-score of 44.9%. It failed to provide a grade or returned an ambiguous grade for 151 samples. Its results showed the same broad split between stronger malignancy scoring and weaker detailed grading.
What the scores do not answer
The study was retrospective and used archived, anonymized prostate-biopsy material; ethics approval was obtained and written consent was waived. For evaluation, the researchers used the same language-model extraction process on original and generated reports to identify malignancy, grading and infiltration attributes. The scores therefore describe those extracted attributes, not a direct assessment by pathologists of the complete generated prose.
The training data came from one institution and one scanner, while the external cohorts used different scanners and had different image distributions. Those differences are an important limitation when interpreting the model’s performance across sites.
Attribute-based scores cannot answer how pathologists would judge the full reports, whether the system would change workload, or whether it could be used for autonomous diagnosis. The study did not provide prospective clinical validation or patient-outcome evidence. Prospective reader studies and direct pathologist review of generated reports are needed to assess those questions.
Paper data and sources
Original title: An end-to-end-trained vision-language model for native-language prostate pathology report generation
Authors: Christian Grashei, Fabian Gülhan, Maximilian Legnar et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-24
DOI: Not available
Original paper · Full text