Cancer Tissue Fraction as a Scanner-Robust Triage Signal for Automated Gleason Grading of Prostate Biopsies: External Validation Across a Middle Eastern Cohort

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear Authors, 

Thank you for your paper entitled "Cancer-Tissue Fraction as a Scanner-Robust Triage Signal for Automated Gleason Grading of Prostate Biopsies: External Validation Across a Middle Eastern Cohort". 

The manuscript addresses an important issue in computational pathology by evaluating the robustness of an automated prostate biopsy grading pipeline across a previously underrepresented population. While automated Gleason grading has already been extensively investigated and validated in predominantly Western cohorts, evidence regarding its performance in Middle Eastern populations remains limited. Therefore, the present study provides useful external validation rather than methodological innovation.

The use of the PAR dataset is particularly valuable because it enables simultaneous assessment of two clinically relevant sources of variability: scanner-dependent image acquisition and inter-pathologist annotation variability. Furthermore, evaluating an independently developed detect-then-grade pipeline, instead of the models originally trained on this dataset, increases the credibility of the reported generalizability.

The introduction is well written and provides a clear overview of the current state of knowledge regarding AI-assisted Gleason grading. The authors appropriately identify the existing evidence, which has predominantly been generated using Western cohorts, and clearly justify the rationale for validating the proposed pipeline in a Middle Eastern population. The study objectives are well defined, particularly the distinction between cancer detection and Gleason grading as two separate clinical tasks. Furthermore, the emphasis on evaluating scanner robustness, including the performance on a compact scanner that may be more accessible in resource-limited settings, highlights the potential clinical relevance of the study. Overall, the introduction logically leads to the study aims and adequately explains the novelty and clinical significance of the work.

However, the novelty of the study should be presented cautiously. The manuscript does not introduce a fundamentally new AI architecture, grading strategy, or biomarker. Instead, its principal contribution lies in demonstrating whether previously established approaches maintain their performance under different geographical, ethnic, laboratory, and hardware conditions.

The methodology is clearly described and sufficiently detailed to ensure transparency and reproducibility. The authors provide an adequate description of the PAR dataset, including its unique characteristics, which enable evaluation of scanner variability and inter-pathologist variability within the same cohort. The sections describing the model architecture, inference pipeline, evaluation strategy, and statistical analysis are well organized and appropriate for the study objectives. The separation of cancer detection from Gleason grading represents a clinically meaningful methodological approach, allowing the performance of each task to be evaluated independently.

The results are presented in a clear, logical, and easy-to-follow manner. The included figures and tables effectively summarize the study findings and facilitate interpretation of the reported performance metrics. The distinction between detection performance and grading performance is consistently maintained throughout the Results section, allowing readers to appreciate the different levels of robustness achieved for these two clinical tasks. The presentation is concise while providing sufficient quantitative information to support the authors' conclusions.

The discussion is balanced, well structured, and appropriately interprets the study findings without overstating their implications. The authors correctly emphasize that this study represents an external validation rather than evidence of a true population effect. They explicitly acknowledge that the simultaneous change in both population and scanning hardware prevents distinguishing whether the observed differences arise from demographic characteristics, laboratory processing, or scanner-related variability. This transparent discussion of the study limitations strengthens the credibility of the manuscript.

The authors also appropriately recognize that the dataset originates from a single Middle Eastern institution, preventing broad generalization to other geographic regions or healthcare systems. Importantly, the practical implications are presented cautiously. The finding that cancer detection generalized well across both scanners and population, whereas grading performance requires local validation before autonomous clinical implementation, represents a realistic and clinically responsible interpretation of the results. Furthermore, the discussion of potential anchoring bias associated with AI-assisted reading demonstrates awareness of the challenges related to integrating artificial intelligence into routine pathology practice.

The concluding statement appropriately summarizes the main findings of the study and remains consistent with the presented results. The authors clearly demonstrate that cancer-tissue fraction represents a robust triage signal for cancer detection across different scanners and that the detect-then-grade pipeline outperformed the ungated approach. The distinction between the robustness of cancer detection and the greater scanner sensitivity of Gleason grading is an important and clinically relevant observation.

The recommendations for implementing a two-stage AI pipeline, performing device-specific validation of the grading component, and confirming performance within the local population are well justified and supported by the findings. Furthermore, highlighting the potential role of low-cost compact scanners adds practical value to the study, particularly for healthcare settings with limited resources.

Overall, the study represents an important validation effort that supports the generalizability of AI-assisted prostate cancer grading across populations, although its scientific contribution is incremental rather than transformative.

Author Response

We thank Reviewer #1 for the thorough and thoughtful evaluation of our manuscript. We appreciate the recognition that this study provides important external validation supporting the generalizability of AI-assisted prostate cancer grading across underrepresented populations. We agree with the characterization of our contribution as a validation effort demonstrating whether established approaches maintain their performance under different geographical, ethnic, laboratory, and hardware conditions, which is how the manuscript frames its novelty. We have incorporated feedback from Reviewers #2 and #3 to strengthen the manuscript as described below.

Reviewer 2 Report

Comments and Suggestions for Authors

This study applies a fixed deep-learning pipeline (PathTools) for prostate biopsy Gleason grading to the PAR dataset (339 slides from 185 patients, Erbil, Iraq) imaged on three scanners (Grundium Ocus40, Hamamatsu NanoZoomer HT 2.0, Leica Aperio GT 450 DX), testing scanner and population transfer simultaneously. The central finding—that cancer-tissue fraction is a scanner-robust detection signal (AUC 0.987–0.991) while grading is scanner-sensitive—is clinically relevant and the detect-then-grade framework is sensible. The manuscript is generally well-written but several issues require substantive revision before acceptance.

 

1. The study reports zero clinical context for the biopsy cohort: no PSA values, no prostate volume, no MRI findings, no biopsy indication (initial vs. repeat vs. active surveillance), no transrectal vs. transperineal route, and no core number per session. Without these data the cohort is a purely pathological abstraction, and a urologist cannot assess whether the 49.6% clinically significant cancer prevalence reflects a referral-biased diagnostic cohort, an active surveillance population, or a mixed screening setting, each of which would carry fundamentally different implications for triage tool generalisability.
2. The ISUP grade distribution under S.A. (7 ISUP 1, 38 ISUP 2, 59 ISUP 3, 30 ISUP 4, 41 ISUP 5) is markedly skewed toward high-grade disease compared with typical Western biopsy cohorts, and ISUP 1 is almost absent (7/339 = 2.1%), which is inconsistent with real-world screening populations where ISUP 1 may constitute 15–25% of cancer diagnoses. This distribution pattern suggests either a highly selected referral cohort or a population-level difference in disease biology, but either way the near-absence of ISUP 1 means the model’s ability to distinguish ISUP 1 from ISUP 2—the single most consequential grading distinction in clinical decision-making for active surveillance vs. definitive treatment—is essentially untested.
3. The 43% inter-slide grade disagreement within the same case (66/154 two-slide cases) is alarmingly high from a clinical perspective and demands urgent discussion of its implications for the worst-grade convention used in patient-level analysis. In urological practice, a case graded ISUP 2 on one slide and ISUP 4 on another would trigger definitive treatment rather than active surveillance; if the reference pathologist’s own readings fluctuate this widely, the patient-level QWK values are not clinically interpretable because the reference standard itself does not represent a stable clinical decision threshold.
4. The clinical utility analysis is incomplete because the authors report only NPV and cleared fraction at the cohort’s 49.6% prevalence without modelling downstream consequences for patient management. A triage tool that auto-clears 50% of biopsies at 98% sensitivity in a high-prevalence cohort would clear a very different subset in a screening population (10–15% prevalence), and the false-negative rate (2% of cancers missed) must be contextualised against the clinical harm of delayed diagnosis in those patients, including missed opportunity for early intervention, psychological burden of delayed repeat biopsy, and potential disease progression during the interval.
5. The study does not report any analysis of the detection failure cases (the 2% of clinically significant cancers missed at 98% sensitivity), which is the clinically most important subset. Characterising these false negatives by grade, tumour extent, and histological features would inform urologists about which cancer types are at highest risk of being auto-cleared and whether additional safeguards (e.g., mandatory second-reader review of all auto-benign slides, or MRI correlation) are needed before clinical deployment.
6. To further strengthen the depth and contextualization of the background and discussion regarding prostate cancer driver mutation genes, biparametric MRI in diagnosis and multi-omics in prognosis, it would be highly beneficial for the authors to discuss several additional studies(PMIDs: 35127694, 40843359, 32609397)in this field.
7. The Definition of “clinically significant prostate cancer” as ISUP ≥2 follows the 2014 ISUP consensus but does not acknowledge ongoing debate about whether ISUP 2 (Gleason 3+4) with <10% pattern 4 should be considered clinically insignificant, which is relevant because the triage tool treats all ISUP ≥2 as positive and could over-triage favourable ISUP 2 cases.
8. The two-slide-per-case design (154 cases with 2 slides, 31 with 1) is unusual for prostate needle biopsy where 10–12 core biopsies are standard; the authors should clarify whether these “slides” represent consolidated core groups (e.g., all left-sided cores on one slide) or individual cores, as this fundamentally affects the clinical interpretability of slide-level metrics.
9. No mention is made of how the AI pipeline handles borderline or ambiguous cases (e.g., atypical small acinar proliferation ASAP, or high-grade prostatic intraepithelial neoplasia HGPIN), which are common in real-world biopsy cohorts and would need to be triaged appropriately in a clinical workflow.

Comments for author File: Comments.pdf

Author Response

1. The study reports zero clinical context for the biopsy cohort: no PSA values, no prostate volume, no MRI findings, no biopsy indication (initial vs. repeat vs. active surveillance), no transrectal vs. transperineal route, and no core number per session. Without these data the cohort is a purely pathological abstraction, and a urologist cannot assess whether the 49.6% clinically significant cancer prevalence reflects a referral-biased diagnostic cohort, an active surveillance population, or a mixed screening setting, each of which would carry fundamentally different implications for triage tool generalisability.

 

We thank the reviewer for this important observation. The PAR dataset [19] comprises consecutive diagnostic prostate needle biopsies from routine clinical practice at PAR Hospital, Erbil, collected between 2013 and 2024. Clinical metadata beyond histopathological grading including PSA values, prostate volume, biopsy indication, and biopsy route were not collected as part of the publicly available dataset, and we acknowledge this as a limitation. The 49.6% clinically significant cancer prevalence is consistent with a hospital-referred diagnostic cohort rather than a screening population. We have added text to the Methods section characterizing the cohort and to the Limitations section acknowledging the absence of clinical metadata.

           

Methods:

“Consecutive diagnostic prostate needle biopsies from routine clinical practice were collected between 2013 and 2024, yielding a hospital-referred diagnostic cohort rather than a screening population.”

 

“Clinical metadata beyond histopathological grading, including PSA values, prostate volume, biopsy indication, and biopsy route, were not collected as part of the publicly available dataset.”

 

Limitations:

“The absence of clinical metadata (PSA, prostate volume, biopsy indication, biopsy route) in the PAR dataset prevents characterization of the cohort beyond its histopathological profile and limits the ability to contextualize model performance against clinical decision-making variables.”

 

1. The ISUP grade distribution under S.A. (7 ISUP 1, 38 ISUP 2, 59 ISUP 3, 30 ISUP 4, 41 ISUP 5) is markedly skewed toward high-grade disease compared with typical Western biopsy cohorts, and ISUP 1 is almost absent (7/339 = 2.1%), which is inconsistent with real-world screening populations where ISUP 1 may constitute 15–25% of cancer diagnoses. This distribution pattern suggests either a highly selected referral cohort or a population-level difference in disease biology, but either way the near-absence of ISUP 1 means the model’s ability to distinguish ISUP 1 from ISUP 2—the single most consequential grading distinction in clinical decision-making for active surveillance vs. definitive treatment—is essentially untested.

 

We agree with the reviewer’s observation. The grade distribution reflects the nature of the PAR cohort as a hospital-referred diagnostic population rather than a screening cohort, which we have now clarified in the manuscript (see previous response). The PAR dataset was selected for its unique property of providing three-scanner digitization of the same physical slides from a non-Western population, a combination unavailable in any other public dataset of which we are aware. The ISUP 1 underrepresentation is inherent to this hospital-referred cohort and limits evaluation of the ISUP 1 versus 2 boundary but does not affect the scanner-robustness and population-transfer questions the dataset was chosen to address. We have added text to the Discussion acknowledging this gap and directing readers to a concurrent investigation that addresses this grading boundary.

 

Limitations:

“The ISUP grade distribution in this cohort is skewed toward high-grade disease, with ISUP 1 constituting only 2.1% of slides, consistent with a hospital-referred diagnostic population rather than a screening cohort where ISUP 1 may represent 15–25% of cancer diagnoses. The near absence of ISUP 1 means that the clinically critical ISUP 1 versus ISUP 2 boundary, a consequential grading distinction for active surveillance eligibility, is essentially untested in the present study. This distinction is a primary focus of a concurrent investigation that evaluates automated grading against disease-free interval outcomes [23].”

 

1. The 43% inter-slide grade disagreement within the same case (66/154 two-slide cases) is alarmingly high from a clinical perspective and demands urgent discussion of its implications for the worst-grade convention used in patient-level analysis. In urological practice, a case graded ISUP 2 on one slide and ISUP 4 on another would trigger definitive treatment rather than active surveillance; if the reference pathologist’s own readings fluctuate this widely, the patient-level QWK values are not clinically interpretable because the reference standard itself does not represent a stable clinical decision threshold.

 

We appreciate the reviewer raising this point. However, we respectfully note that the 43% inter-slide discordance reflects the spatial heterogeneity of tumor grade across anatomically distinct biopsy sites rather than instability of the reference standard. In the PAR dataset, each slide represents the consolidated tissue from a separate collection container corresponding to a distinct anatomical biopsy site (e.g., six left-sided cores versus six right-sided cores) [19]. Prostate cancer is well established to harbor heterogeneous Gleason patterns across different regions of the gland, and a patient whose left-sided cores show ISUP 2 while right-sided cores show ISUP 4 carries both grades in different anatomical locations. The worst-grade convention applied here, assigning a case the highest ISUP grade across its slides, is standard clinical reporting practice [2, 3] and yields the treatment-determining grade a urologist would derive from a complete biopsy report. In the reviewer’s own example, an ISUP 2 on one slide and ISUP 4 on another would correctly result in a case-level ISUP 4, which is precisely the grade that triggers definitive treatment. We have added a paragraph to the Discussion clarifying this point.

 

Discussion:

“The 43% inter-slide grade discordance within two-slide cases (66 of 154) reflects spatial heterogeneity of tumor grade across anatomically distinct biopsy sites rather than instability of the reference standard. Each slide in the PAR dataset represents tissue from a separate collection container corresponding to a distinct anatomical site [19], and prostate cancer commonly harbors heterogeneous Gleason patterns across the gland. The worst-grade convention applied here, assigning a case the highest ISUP across its slides, is standard clinical practice [2, 3] and yields the same treatment-determining grade a urologist would derive from a complete biopsy report.”

 

1. The clinical utility analysis is incomplete because the authors report only NPV and cleared fraction at the cohort’s 49.6% prevalence without modelling downstream consequences for patient management. A triage tool that auto-clears 50% of biopsies at 98% sensitivity in a high-prevalence cohort would clear a very different subset in a screening population (10–15% prevalence), and the false-negative rate (2% of cancers missed) must be contextualised against the clinical harm of delayed diagnosis in those patients, including missed opportunity for early intervention, psychological burden of delayed repeat biopsy, and potential disease progression during the interval.

 

We agree that the prevalence-dependence of the triage operating characteristics warrants more explicit discussion. The manuscript already notes that the cleared fraction is prevalence-dependent, but we have expanded this passage to address the screening-prevalence scenario and to contextualize the clinical consequences of false-negative results.

 

Discussion:

'The magnitude of that saving is prevalence-dependent: in this cohort, which is roughly half clinically significant cancer, 45 to 48% of slides are auto-cleared at 98% sensitivity. At a screening-population prevalence of 10–15%, the same sensitivity threshold would yield an NPV exceeding 99.5%, and the auto-cleared fraction would increase substantially. The false-negative rate, approximately 2% of clinically significant cancers, must be contextualized against the clinical consequences of delayed diagnosis, including missed opportunity for early intervention and potential disease progression during the interval to repeat biopsy. We acknowledge that prospective evaluation of downstream patient management consequences, including the psychological burden of delayed diagnosis and the adequacy of safety nets such as mandatory second-reader review or MRI correlation for auto-benign cases, is required before autonomous deployment in any clinical setting.”

 

1. The study does not report any analysis of the detection failure cases (the 2% of clinically significant cancers missed at 98% sensitivity), which is the clinically most important subset. Characterising these false negatives by grade, tumour extent, and histological features would inform urologists about which cancer types are at highest risk of being auto-cleared and whether additional safeguards (e.g., mandatory second-reader review of all auto-benign slides, or MRI correlation) are needed before clinical deployment.

 

We agree that a detailed analysis of failure cases would be important prior to any clinical deployment, especially when running prospective analysis. Here, the focus is on retrospective analysis of a dataset that already has been graded by 3 separate pathologists (see original PRA publication).  

 

1. To further strengthen the depth and contextualization of the background and discussion regarding prostate cancer driver mutation genes, biparametric MRI in diagnosis and multi-omics in prognosis, it would be highly beneficial for the authors to discuss several additional studies (PMIDs: 35127694, 40843359, 32609397) in this field.

 

We thank the reviewer for these references. While the present study focuses specifically on histopathological grading, we agree that situating our work within the broader prostate cancer diagnostic landscape adds valuable context. We have added a sentence to the Introduction acknowledging these complementary diagnostic modalities and citing the suggested studies.

 

Introduction:

“Beyond histopathological grading, prostate cancer diagnosis and prognosis increasingly integrate complementary modalities including biparametric MRI combined with PSA-related indicators [6], genomic characterization of aneuploidy driver genes [7], and multi-omics profiling of the tumor immune microenvironment [8]; the present study addresses the histopathology component of this evolving diagnostic ecosystem.”

 

1. The Definition of “clinically significant prostate cancer” as ISUP ≥2 follows the 2014 ISUP consensus but does not acknowledge ongoing debate about whether ISUP 2 (Gleason 3+4) with <10% pattern 4 should be considered clinically insignificant, which is relevant because the triage tool treats all ISUP ≥2 as positive and could over-triage favourable ISUP 2 cases.

 

We agree and have added a sentence acknowledging this ongoing debate.

 

Methods:

We define clinically significant prostate cancer as ISUP grade group ≥2, following the 2014 ISUP consensus [2]. We note that whether ISUP 2 (Gleason 3+4) cases with less than 10% pattern 4 should be reclassified as clinically insignificant remains an active area of debate; the present triage tool treats all ISUP ≥2 as positive and may therefore over-triage a minority of favorable ISUP 2 cases, a consideration for prospective clinical evaluation.

 

1. The two-slide-per-case design (154 cases with 2 slides, 31 with 1) is unusual for prostate needle biopsy where 10–12 core biopsies are standard; the authors should clarify whether these “slides” represent consolidated core groups (e.g., all left-sided cores on one slide) or individual cores, as this fundamentally affects the clinical interpretability of slide-level metrics.

 

We thank the reviewer for requesting this clarification. In the PAR dataset, each slide represents the consolidated tissue from a single collection container corresponding to an anatomical biopsy site. As described in the primary dataset publication [19], when biopsies from each side were submitted in separate containers, separate slides were generated; when biopsies from both sides were submitted in a single container, they were combined on a single slide. These slides therefore represent anatomically grouped core sets, not individual cores. We have added this clarification to the Methods.

Methods, Section 2.1:

“Each patient contributed one or two slides, where each slide represents the consolidated tissue from a single collection container corresponding to an anatomical biopsy site (e.g., left-sided or right-sided cores); when biopsies from both sides were submitted in a single container, they were combined on one slide [19]. These slides therefore represent consolidated core groups rather than individual cores.”

 

1. No mention is made of how the AI pipeline handles borderline or ambiguous cases (e.g., atypical small acinar proliferation ASAP, or high-grade prostatic intraepithelial neoplasia HGPIN), which are common in real-world biopsy cohorts and would need to be triaged appropriately in a clinical workflow.

 

We thank the reviewer for this clarification that has now been added to the limitation section:

 

The pipeline was not evaluated on ASAP or HGPIN, entities likely routed to the benign class; real-world deployment would require explicit handling, potentially via mandatory pathologist review of all auto-benign slides.

Reviewer 3 Report

Comments and Suggestions for Authors

Thank you for submitting this manuscript entitled 'Cancer-Tissue Fraction as a Scanner-Robust Triage Signal for Automated Gleason Grading of Prostate Biopsies: External Validation Across a Middle Eastern Cohort'. I have some comments:

• The study relies on one reference pathologist, which may introduce bias. Consider using consensus grading by multiple expert uropathologists in future studies.
• More details on the PathTools training dataset (number of cases, scanner types, grade distribution, and prior validations) would help readers assess generalizability.
• The term "scanner-robust" may be too strong. Consider stating that no significant performance differences were observed between scanners in this dataset.
• As this is a single-center study, the findings should be interpreted cautiously. Additional multicenter studies are needed before broader conclusions can be drawn.
• Since some authors are affiliated with the model developer, the manuscript should emphasize the need for independent external validation.
• Please include more technical information about the scanners, such as resolution, pixel size, magnification and image-processing settings.
• Clarify how differences between pathologists were resolved and whether alternative reference standards were evaluated.
• Please expand the discussion on why Hamamatsu scans were more often graded as ISUP 5 and discuss possible technical causes.
• Please report the number of cancer-positive cases included in the grading analyses to improve clarity.
• Shortening repetitive sections, particularly comparisons with the PAR study, would improve readability and focus.

Author Response

Thank you for submitting this manuscript entitled 'Cancer-Tissue Fraction as a Scanner-Robust Triage Signal for Automated Gleason Grading of Prostate Biopsies: External Validation Across a Middle Eastern Cohort'. I have some comments:

• The study relies on one reference pathologist, which may introduce bias. Consider using consensus grading by multiple expert uropathologists in future studies.

 

We acknowledge this limitation. In the present study, pathologist S.A. served as the primary reference reader because S.A. graded the full cohort. To mitigate single-reader bias, we validated S.A.’s grades against a subspecialist uropathologist (QWK 0.84, 95% CI 0.70–0.93 on the 59-slide overlap) and reported sensitivity analyses against an independent second full-cohort reader (H.M.) as well as majority-vote and highest-grade reference rules (Table 3). We agree that consensus grading by multiple expert uropathologists should be incorporated in future multi-site validation studies and have added a note to the Limitations section.

 

Limitations:

“The reference standard is single-reader-dominant: no consensus label or pixel-level annotation exists for PAR, the two full-cohort readers agree only moderately, and scoring the model against S.A. while comparing to an S.A.-based ceiling shares a common reader. We mitigate the last point by also reporting agreement against H.M., which was similar, but it is not fully removed. Future validation studies should incorporate consensus grading by multiple expert uropathologists.”

 

• More details on the PathTools training dataset (number of cases, scanner types, grade distribution, and prior validations) would help readers assess generalizability.

 

We thank the reviewer for this comment. PathTools offered a black box pre-trained algorithm for this study. All architecture and training details comprise intellectual property of the company and are secondary to the claims and tests performed in this study.  

 

• The term "scanner-robust" may be too strong. Consider stating that no significant performance differences were observed between scanners in this dataset.

 

We appreciate this suggestion. We have qualified the language in the body of the manuscript where appropriate, noting that no statistically significant between-scanner differences in cancer-tissue fraction were observed in this dataset. We retain ‘scanner-robust’ in the title as it refers specifically to the triage signal (cancer-tissue fraction), which showed statistical equivalence across all three scanners (paired AUC differences ≤0.003), while the grading component is explicitly described as scanner-sensitive throughout the manuscript.

 

• As this is a single-center study, the findings should be interpreted cautiously. Additional multicenter studies are needed before broader conclusions can be drawn.

 

We agree with the reviewer’s point. The Limitations section already acknowledges the single-center design. We have strengthened this language to explicitly state the need for multi-center validation.

 

Limitations:

“As a single-center study, these findings should be interpreted cautiously; multi-center validation across diverse institutions, populations, and laboratory protocols is required before broader conclusions about the pipeline’s generalizability can be drawn.”

 

• Since some authors are affiliated with the model developer, the manuscript should emphasize the need for independent external validation.

 

We have added a sentence to the Discussion regarding this point.

 

Discussion:

“We note that some authors are affiliated with the developers of the PathTools pipeline. Independent external validation by groups without developer involvement will therefore be an important component of future studies.”

 

• Please include more technical information about the scanners, such as resolution, pixel size, magnification and image-processing settings.

 

We have added scanner technical specifications to the Methods section. These include native scanning magnification, pixel size (μm/pixel), and image format.

 

Methods:

“Each slide was scanned on three devices: a Grundium Ocus40 (compact portable scanner; 40× magnification; 0.25 µm/pixel; .svs format), Hamamatsu NanoZoomer HT 2.0 (mid-tier clinical scanner; 40× magnification; 0.22 µm/pixel, .ndpi format), and Leica Aperio GT 450 DX (clinical flagship; 40× magnification; 0.26 µm/pixel, .svs format), giving 1017 whole-slide images. Slides were produced with standard laboratory scanner software and WSIs were stored in the scanners’ proprietary formats.”

 

• Clarify how differences between pathologists were resolved and whether alternative reference standards were evaluated.

 

Pathologist differences were not resolved by adjudication; instead, each reader’s grades were used as independent reference standards. Pathologist S.A. served as the primary reference because S.A. graded the full cohort, while results against H.M. and against majority-vote and highest-grade reference rules are reported as sensitivity analyses in Table 3. This design was deliberate: the study aims to assess model performance under real-world inter-observer variability rather than against an artificial consensus. We have clarified this in the Methods.

 

Methods:

“Pathologist grades were not adjudicated to consensus; instead, each reader’s independent grades served as a separate reference standard. This design preserves the inter-observer variability inherent in clinical grading and allows assessment of model performance under realistic reference conditions.”

 

• Please expand the discussion on why Hamamatsu scans were more often graded as ISUP 5 and discuss possible technical causes.

 

We have expanded the discussion of the Hamamatsu upward grade shift to include possible technical explanations.

 

Discussion:

“Grading of the detected cancers is where scanner choice matters and should be validated per device. The mid-tier Hamamatsu showed a directional Gleason-5 bias that is a candidate for stain normalization or grade recalibration. Possible technical explanations include differences in color reproduction (Hamamatsu’s color response may render pattern 4 tissue with staining intensity or morphological characteristics that the model associates with pattern 5), compression artefacts, or focus-plane differences that alter perceived nuclear features used by the model for grade assignment. The directional shift may also be partially attributable to a reference effect, since the second reader H.M. also assigned ISUP 5 more often than S.A. (86 versus 41 slides).”

 

• Please report the number of cancer-positive cases included in the grading analyses to improve clarity.

 

We have added explicit counts of cancer-positive slides to the grading results sections for clarity.

 

Results:

“Grading agreement was evaluated on the 175 pathologist-confirmed cancer slides (ISUP ≥1 under S.A.), of which 168 were clinically significant (ISUP ≥2).”

 

• Shortening repetitive sections, particularly comparisons with the PAR study, would improve readability and focus.

 

We agree and have trimmed the Discussion to reduce redundant comparisons with the original PAR study findings. Passages that restate PAR results without adding interpretive value have been condensed or removed.

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Authors have addressed all comments provided by reviewers and now it is acceptable for publication in current version.