Quantitative Texture Analysis of Cervical Cytology Identifies Endometrial Lesions in Atypical Glandular Cells on Liquid-Based Cytology: A Pilot Study
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis study presents a well-designed and clinically relevant approach to distinguishing endometrial lesions among atypical glandular cell cases using quantitative analysis of liquid-based cervical cytology. The methodology is clearly described, technically sound, and appropriately integrates image-derived features with machine-learning models that can be readily incorporated into existing cytopathology workflows.
1. The authors should more explicitly acknowledge the potential confounding effect introduced by the temporally separated use of SurePath and ThinPrep platforms and briefly discuss how this may have influenced comparative performance.
2. The limitations related to the absence of external validation should be stated more clearly, emphasizing that the reported performance reflects internal cross-validation only.
3. A short clarification regarding the clinical context in which this quantitative approach would be most useful (e.g., AGC cases with subtle morphology or HPV-negative results) would enhance the practical interpretation of the findings.
With these minor revisions, the manuscript would be suitable for publication.
Author Response
Dear Editor,
We thank you and the reviewers for the careful evaluation of our manuscript entitled “Quantitative Texture Analysis of Cervical Cytology Identifies Endometrial Lesions in Atypical Glandular Cells on Liquid-based Cytology: A Pilot Study” (Manuscript ID: diagnostics-4115927). We have revised the manuscript in response to all comments, and we believe the revised version has been improved in clarity, transparency, and clinical interpretation. In addition, we identified an inadvertent calculation error affecting Figure 1 during the revision process. We have corrected the calculations, updated the corresponding numerical values in the main text, and replaced Figure 1 with the corrected version. This correction does not affect the main analyses or the overall conclusions of the study. We appreciate the opportunity to revise our work and hope that the revised manuscript is now suitable for publication in Diagnostics.
Sincerely,
Toshimichi Onuma
Department of Obstetrics and Gynecology, Faculty of Medical Sciences, University of Fukui, 23-3 Matsuoka-Shimoaizuki, Eiheiji-cho, Fukui 910-1193, Japan
Tel: +81-776-61-8431
Fax: +81-776-61-8170
E-mail: toonuma@u-fukui.ac.jp
#Reviewer 1
This study presents a well-designed and clinically relevant approach to distinguishing endometrial lesions among atypical glandular cell cases using quantitative analysis of liquid-based cervical cytology. The methodology is clearly described, technically sound, and appropriately integrates image-derived features with machine-learning models that can be readily incorporated into existing cytopathology workflows.
Response: Thank you for your careful review and constructive comments. We have revised the manuscript accordingly, as detailed below.
- The authors should more explicitly acknowledge the potential confounding effect introduced by the temporally separated use of SurePath and ThinPrep platforms and briefly discuss how this may have influenced comparative performance.
Response: Thank you for this important comment. We agree that the temporally separated use of SurePath and ThinPrep may introduce time related confounding and could influence comparative performance. We have acknowledged this point in the Limitations section, and we clarified that our analyses were intended primarily as within platform evaluations and that cross platform comparisons should be interpreted cautiously.
Line 316
Textural features may be influenced by pre-analytical and analytical factors, including staining procedures, scanner calibration, and platform architecture. Because SurePath and ThinPrep were used in temporally separated periods, cross-platform differences may be confounded by changes in case mix. Therefore, our analyses were intended primarily as within-platform evaluations, and any cross-platform comparisons should be interpreted cautiously. Importantly, the inclusion of both SurePath and ThinPrep helps clarify preparation-dependent behavior and indicates the need for standardization. Thus, prospective, multicenter validation with standardized workflows will be necessary.
- The limitations related to the absence of external validation should be stated more clearly, emphasizing that the reported performance reflects internal cross-validation only.
Response: We appreciate this suggestion and agree that this limitation should be stated more clearly. We revised the Limitations section to show that external validation was not performed and that the reported performance reflects internal cross-validation only. We also indicate the need for external validation in independent multi-site cohorts.
Line 308
This study had some limitations. It was a single-center retrospective study without external validation, and the reported performance reflects internal cross-validation only. Therefore, external multi-site validation is required.
- A short clarification regarding the clinical context in which this quantitative approach would be most useful (e.g., AGC cases with subtle morphology or HPV-negative results) would enhance the practical interpretation of the findings.
Response: Thank you for this helpful recommendation. We agree that specifying the clinical context improves practical interpretation. We have added a brief statement in the Discussion to clarify where this quantitative approach may be most useful, particularly as decision support in AGC cases with subtle or equivocal morphology and as a potential aid to identifying AEH or EC that might otherwise be missed in HPV negative cases within HPV primary screening workflows.
Line 305
This approach may be most useful as a decision support adjunct in AGC cases with subtle or equivocal morphology, and it may help identify AEH/EC that could otherwise be missed in HPV negative cases under HPV primary screening.
Reviewer 2 Report
Comments and Suggestions for AuthorsThis is a well-written study with original data worthy of publication after major revision according to the following comments.
1) Title: "Sure Path and ThinPrep" should be better changed to "Liquid-based Cytology".
2) Lines 48-49: “Previous studies have shown…” should be changed to “A previous study has shown…”; the country where this study was conducted should be added.
3) Line 53 : “400,000 new cases reported annually” should be changed to “400,000 new cases worldwide reported annually”.
4) Line 69: The expression "naked eye" is not accurate; Please change "naked eye" to e.g. "unaided microscopy" or "classical microscopy".
5) Line 127: All acronyms used in Figure 1 should be explained in the figure legend.
6) Lines 136-139, lines 159-161 and lines 246-248: These acronyms and their explanations do not fit here. Acronyms should be explained (i.e. written in full) upon their first appearance in the text. The full list of acronyms with their explanations at the end of the main text should be reviewed and any missing acronyms should be added, if this is needed.
7) Line 250: The two liquid-based cytology methods should be mentioned.
8) Discussion: The authors should discuss their findings and conclusions in comparison to those of other authors previously published in the literature.
9) Lines 294-301: In this paragraph the authors should add the strengths of their study (e.g. originality).
Author Response
Dear Editor,
We thank you and the reviewers for the careful evaluation of our manuscript entitled “Quantitative Texture Analysis of Cervical Cytology Identifies Endometrial Lesions in Atypical Glandular Cells on Liquid-based Cytology: A Pilot Study” (Manuscript ID: diagnostics-4115927). We have revised the manuscript in response to all comments, and we believe the revised version has been improved in clarity, transparency, and clinical interpretation. In addition, we identified an inadvertent calculation error affecting Figure 1 during the revision process. We have corrected the calculations, updated the corresponding numerical values in the main text, and replaced Figure 1 with the corrected version. This correction does not affect the main analyses or the overall conclusions of the study. We appreciate the opportunity to revise our work and hope that the revised manuscript is now suitable for publication in Diagnostics.
Sincerely,
Toshimichi Onuma
Department of Obstetrics and Gynecology, Faculty of Medical Sciences, University of Fukui, 23-3 Matsuoka-Shimoaizuki, Eiheiji-cho, Fukui 910-1193, Japan
Tel: +81-776-61-8431
Fax: +81-776-61-8170
E-mail: toonuma@u-fukui.ac.jp
#Reviewer 2
This is a well-written study with original data worthy of publication after major revision according to the following comments.
Thank you for your careful review and constructive comments. We have revised the manuscript accordingly, as detailed below.
1) Title: "Sure Path and ThinPrep" should be better changed to "Liquid-based Cytology".
Response: Thank you for this valuable suggestion. We agree that the title should be more generally to reflect liquid-based cytology rather than specific platforms. We have revised the title as follows.
Quantitative Texture Analysis of Cervical Cytology Identifies Endometrial Lesions in Atypical Glandular Cells on Liquid-based Cytology: A Pilot Study
2) Lines 48-49: “Previous studies have shown…” should be changed to “A previous study has shown…”; the country where this study was conducted should be added.
Response: Thank you for pointing this out. We have revised the sentence.
Line 48
A previous study conducted in the United States has shown that approximately 5.2% of AGC cases are associated with malignancy.
3) Line 53 : “400,000 new cases reported annually” should be changed to “400,000 new cases worldwide reported annually”.
Response: Thank you for this correction. We have revised the sentence.
Line 53
Recent global estimates and trend analyses show substantial growth in EC burden since the 1990s, with 400,000 new cases worldwide reported annually in 2020–2022
4) Line 69: The expression "naked eye" is not accurate; Please change "naked eye" to e.g. "unaided microscopy" or "classical microscopy".
Response: Thank you for this helpful suggestion. We agree that “naked eye” is inappropriate in this context and replaced it with “unaided microscopy.”
Line 68
These computational methods can complement traditional, experience-dependent assessments through revealing subtle differences not readily appreciable to an unaided microscopy
5) Line 127: All acronyms used in Figure 1 should be explained in the figure legend.
Response: Thank you for your careful review, which prompted us to re-check Figure 1. We agree that all acronyms appearing in Figure 1 should be defined in the figure legend.
During revision, we noticed that an incorrect calculation had inadvertently been used in preparing Figure 1. We re-checked the underlying data, corrected the calculations, replaced Figure 1 with the corrected version, and updated the corresponding numerical values in the main text accordingly. We deeply appreciate your comment.
Line 77
Between 2014 and 2024, we identified 161 cases of AGCs at the University of Fukui Hospital in which liquid-based cervical cytology (LBC; SurePath or ThinPrep) had been used.
Line 96
Figure 1 shows the pathological results of the AGC cases, of which 36.6% (59/161) were lesions of endometrial origin, and 37.9% (61/161) were classified as normal.
Line 106
Figure 1. Pathological outcomes of cases with AGC on cervical cytology Final diagnoses were categorized as normal (n = 61) or AEH/EC (n = 53). ADC, adenocarcinoma; AEH/EC, atypical en-dometrial hyperplasia or endometrial cancer; AGC, atypical glandular cells; AIS, adenocarcinoma in situ; BOT, borderline ovarian tumor; CIN1, cervical intraepithelial neoplasia 1; CIN2, cervical intraepithelial neoplasia 2; CIN3, cervical intraepithelial neoplasia 3; OVCA, ovarian cancer; SCC, squamous cell carcinoma
6) Lines 136-139, lines 159-161 and lines 246-248: These acronyms and their explanations do not fit here. Acronyms should be explained (i.e. written in full) upon their first appearance in the text. The full list of acronyms with their explanations at the end of the main text should be reviewed and any missing acronyms should be added, if this is needed.
Response: Thank you for this helpful comment. We agree that acronyms should be defined at their first appearance in the main text and that the placement of acronym definitions should be clear and consistent. We have now revised the manuscript to improve readability and to ensure that acronym usage and definitions follow standard conventions. In addition we corrected the titles of Tables 1 and 2 to more accurately reflect their contents. We also added a brief footnote to Tables 1 and 2 noting that, in some cases, multiple thresholds yielded identical maximum Youden index values. We removed “WSI” from the Abbreviations list because it is not used in the main text. We also revised Figure 4 so that the CV AUC is reported to three decimal places for consistency throughout the manuscript.
Line 77
Between 2014 and 2024, we identified 161 cases of AGCs at the University of Fukui Hospital in which liquid-based cervical cytology (LBC; SurePath or ThinPrep) had been used. The AGC subcategories were not routinely assigned [23]. The platform varied according to period—SurePath (2014–2017), ThinPrep (2018–2023), and SurePath again in 2024—reflecting transitions prompted by institutional equipment upgrades. Reference diagnoses were determined from histologic assessment or follow-up cytology. Histological diagnosis was derived from biopsy or surgical specimens. The benign (normal) category included the following: (i) cases with benign histologic findings and no evidence of atypical endometrial hyperplasia or endometrial cancer (AEH/EC) and/or (ii) cases in which repeat cervical cytology during follow-up was negative for intraepithelial lesion or malignancy without a subsequent clinical diagnosis of endometrial neoplasia. Trained pathology specialists performed all the histopathologic and cytologic assessments. In this study, AEH/EC comprised atypical endometrial hyperplasia and endometrial cancer, including endometrioid carcinoma and non-endometrioid histologies such as serous and clear cell carcinoma.
Line 132
Figure 2. Workflow for quantitative analysis of AGC- classified liquid-based cervical cytology. Whole-slide images from LBC (SurePath or ThinPrep) were digitized and preprocessed. Cell clus-ters were automatically detected on downsampled images and mapped back to full-resolution co-ordinates. For each detected cluster, quantitative readouts were extracted from regions of interest (texture, intensity, and geometric descriptors). Case-level data were then generated through summarizing cluster-wise measurements (cluster count and distributional statistics: mean, SD, minimum, maximum, median, Q75, and Q95). These aggregated outputs formed the inputs for subse-quent diagnostic analyses. GLCM, gray- level co-occurrence matrix; HSB, hue, saturation, bright-ness; LBC, liquid-based cytology; LBP, local binary pattern; max, maximum; min, minimum; OD, optical density; Q75, 75th percentile; Q95, 95th percentile; RGB, red, green, blue; SD, standard deviation.
Line 167
The final feature subset and hyperparameters were selected by maximizing performance under stratified 5-fold cross-validation (CV) [33].
Line 185
Table 1. Highest 10 ROC curve analysis results between the normal and AEH/EC lesion groups using SurePath.
Line 187
AEH/EC, atypical endometrial hyperplasia or endometrial cancer; AUC, area under the curve; OD, optical density; max., maximum; min., minimum; ROC, receiver operating characteristic; SD, standard deviation. * In some cases, multiple thresholds yielded identical maximum Youden in-dex values.
Line 200
Using the locked three-feature subset (minimum of hematoxylin, SD; SD of hema-toxylin, SD; and minimum of red, SD), and tuned RF hyperparameters, a stratified 5-fold cross-validation yielded a CV AUC of 0.805 (95% CI, 0.683–0.927). At the Youden threshold (0.471), sensitivity was 0.840 and specificity was 0.811 (Figure 3).
Line 206
Figure 3. ROC curve for AGC cases on SurePath with 5-fold cross-validation. ROC curve analysis for distinguishing AEH/EC from normal within AGC cases on SurePath cytology. Performance was estimated using stratified 5-fold cross-validation with out-of-fold predictions. The CV AUC was 0.805. AEH/EC, atypical endometrial hyperplasia or endometrial cancer; AGC, atypical glandular cells; CV AUC, cross-validated area under the curve; ROC, receiver operating characteristic.
Line 226
Table 2. Highest 10 results of ROC curve analysis between the normal and AEH/EC lesion groups using ThinPrep.
Line 228
AEH/EC, atypical endometrial hyperplasia or endometrial cancer; AUC, area under the curve; OD, optical density; max., maximum; min., minimum; ROC, receiver operating characteristic; SD, standard deviation. * In some cases, multiple thresholds yielded identical maximum Youden index values.
Line 241
Using the locked five-feature subset (SD of brightness, median; minimum of green, minimum; minimum of hematoxylin, median; mean of length; and minimum of sum of squares variances) and tuned RF hyperparameters, a stratified 5-fold cross-validation yielded a CV AUC of 0.887 (95% CI, 0.787–0.987). At the Youden threshold (0.515), sensitivity was 0.821 and specificity was 0.958 (Figure 4).
Line 248
Figure 4. ROC curve for AGC cases on ThinPrep with 5-fold cross-validation. ROC curve analysis for distinguishing AEH/EC from normal within AGC cases on ThinPrep cytology. Performance was estimated using stratified 5-fold cross-validation with out-of-fold predictions. The CV AUC was 0.887. AEH/EC, atypical endometrial hyperplasia or endometrial cancer; AGC, atypical glandular cells; CV AUC, cross-validated area under the curve; ROC, receiver operating characteristic
7) Line 250: The two liquid-based cytology methods should be mentioned.
Thank you for this comment. We revised the sentence as follows.
Line 255
Across SurePath and ThinPrep liquid-based preparations, univariate ROC analyses showed only moderate discrimination between normal cases and AEH/EC cases overall but consistently stronger separation post-menopause.
8) Discussion: The authors should discuss their findings and conclusions in comparison to those of other authors previously published in the literature.
Response: Thank you for this important suggestion. We agree that our findings and conclusions should be discussed in comparison to previously published literature. We revised the Discussion to compare our results with prior reports on the clinical significance and heterogeneity of AGC, the association of endometrial cell findings with endometrial pathology in older or postmenopausal women, interobserver variability in AGC interpretation, and the feasibility and validation of AI-assisted cytology.
Line 265
Across both preparations, AEH/EC could be differentiated from normal cases. Univariate ROC analyses indicated moderate separability overall with stronger separa-tion post-menopause, and compact RF models further improved performance. This is consistent with prior clinical literature indicating that a subset of AGC shows cytologic features suggestive of AEH/EC and is associated with a clinically meaningful risk of underlying neoplasia [4–6].The stronger separation observed after menopause is clini-cally plausible, as reports have shown that endometrial cell findings on cervical cytology in older or postmenopausal women are associated with EC [10,11]. These findings suggest that quantitative image features can assist where human visual assessment is challenging, offering operational leverage, particularly given the interobserver variability in AGC interpretation [15,16]. RGB-derived values should not be equated with pure hematoxy-lin/eosin concentrations without stain unmixing or normalization [35]. Preparation- and scanner-related variability also indicates the need for harmonization in future external validation [36]. If these limitations can be overcome, image analysis may extend diag-nostic capability in selected contexts.
line 293
Image-analysis-based adjuncts are technically feasible and increasingly validated. AI–assisted or computer-aided cytology can triage slides at scale, improving sensitivity while maintaining clinically acceptable specificity in both population screening and prospective evaluations [41–44]. These software-based methods analyze quantitative morphology, intensity, and texture features from standard liquid-based preparations. These image-based measures offer an objective and reproducible aid to routine review. Because primary HPV screening is designed to detect cervical diseases, endometrial le-sions such as AEH/EC fall outside the intended detection scope of HPV testing and may therefore be overlooked, even though cervicovaginal cytology has historically identified a substantial proportion of AEH/EC incidentally [2,45]. In this context, our results extend prior AI-cytology work that has largely focused on cervical lesion detection and grading [22,41–44] by evaluating AEH/EC discrimination within AGC, a clinically difficult di-agnostic setting.
This approach may be most useful as a decision support adjunct in AGC cases with subtle or equivocal morphology, and it may help identify AEH/EC that could otherwise be missed in HPV negative cases under HPV primary screening.
9) Lines 294-301: In this paragraph the authors should add the strengths of their study (e.g. originality).
Response: Thank you for this helpful comment. We agree that, in addition to describing the limitations, the strengths and originality of this pilot study should be stated more clearly. We revised the Limitations paragraph to highlight the practical relevance and novelty of our approach while maintaining a balanced discussion of study limitations.
Line 308
This study had some limitations. It was a single-center retrospective study without external validation, and the reported performance reflects internal cross-validation only. Therefore, external multi-site validation is required. However, the cohort reflects a re-al-world clinical workflow in which liquid-based cervical cytology cases classified as AGC by trained specialists are managed in routine practice, supporting the practical relevance of the findings. Owing to the limited number of cases, RF-based analyses were not performed within the post-menopausal subgroup. Nevertheless, the consistently stronger discrimination observed after menopause suggests a clinically relevant subgroup and provides a clear rationale for targeted validation. Textural features may be influenced by pre-analytical and analytical factors, including staining procedures, scanner calibration, and platform architecture. Because SurePath and ThinPrep were used in temporally separated periods, cross-platform differences may be confounded by changes in case mix. Therefore, our analyses were intended primarily as within-platform evaluations, and any cross-platform comparisons should be interpreted cautiously. Importantly, the inclusion of both SurePath and ThinPrep helps clarify preparation-dependent behavior and indicates the need for standardization. Thus, prospective, multicenter validation with standardized workflows will be necessary.
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors have responded adequately to previous reviewer comments and now this paper may be accepted for publication in its present form.
