Review Reports
- Ezgi Katı 1,
- Gizem Eğilmez Özkılıç 1 and
- İbrahim Şevki Bayrakdar 3
- et al.
Reviewer 1: Anonymous Reviewer 2: Anonymous Reviewer 3: Anonymous
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsIn this study, the researcher used nnU-Net V2 for pulp stone segmentation on 3D cone-beam computed tomography (CBCT) images. In this study, they used 90 CBCT examinations and created different VOIs (238 different VOI volumes) from these images. The reported results, such as DICE=0.183, are very weak for segmentation performance. The author should train the model more carefully for improving the segmentation performance. My listed comments can be found below:
1- In the manuscript, you have mentioned ‘ROIs,’ while you have used volume of interest (VOI), and you should correct this point in the manuscript.
2- In this study, 3 different experts have been used for data labeling and masking. You should provide interobserver agreement between these experts. (no change needed)
3- The reported segmentation performance metrics for full-volume (Dice similarity coefficient of 0.183 and a Jaccard index of 0.119, precision of 0.254, and recall of 0.205) and ROI-based model (Table 2) are very bad. You should find the cause of this observation and solve it for model performance improvement.
4- Please provide high-res figures with increased text font size for better visualization.
5- According to the ROC curve in Figure 4, the trained model has very high specificity and low sensitivity. Please provide a scientific reason for this observation.
6- Please provide images of the predictions alongside the ground-truth for comparison with different test datasets.
Author Response
Comment 1- In the manuscript, you have mentioned ‘ROIs,’ while you have used volume of interest (VOI), and you should correct this point in the manuscript.
Response: We appreciate the reviewer’s clarification regarding the appropriate three-dimensional terminology. As the cropped areas used in this study represent volumetric tooth data, all remaining instances of “region of interest” have been replaced with “volume of interest.” The abbreviation “VOI” has also been defined accordingly in the abbreviation list.
Page 1, lines 14–17; page 5, lines 168–172; page 6, lines 195–198; page 19, lines 653–655.
Comment 2- In this study, 3 different experts have been used for data labeling and masking. You should provide interobserver agreement between these experts.
Response: The reviewer raises an important point regarding the reproducibility of the reference annotations. In the present study, both the confirmation of pulp-stone presence and the evaluation of voxel-wise mask boundaries were performed through a joint consensus procedure. Separate observer-specific assessments and segmentation masks were not generated; therefore, quantitative interobserver agreement could not be calculated retrospectively. We have clarified the consensus procedure in the Methods and explicitly acknowledged the absence of a quantitative interobserver agreement assessment as a limitation of the study.
Page 5, lines 185–190; page 6, lines 199–206; page 17, lines 575–577.
Comment 3- The reported segmentation performance metrics for full-volume (Dice similarity coefficient of 0.183 and a Jaccard index of 0.119, precision of 0.254, and recall of 0.205) and ROI-based model (Table 2) are very bad. You should find the cause of this observation and solve it for model performance improvement.
Response: The reviewer raises a valid concern regarding the observed segmentation performance. We agree that the full-volume results indicate poor performance and that the VOI-based results represent only moderate segmentation performance. Additional model retraining or architectural optimization was not performed during the present revision; therefore, we do not claim that the performance limitation has been resolved. Instead, we revised the Results, Discussion, Limitations, and Conclusions to provide a more cautious interpretation and to discuss the most plausible causes of the observed performance, including severe voxel-level class imbalance, the small size of pulp stones, partial-volume effects, CBCT image noise, indistinct lesion boundaries, and radiodensity similarities with adjacent structures.
Specifically, we now characterize the full-volume performance as poor and the VOI-based performance as moderate, clarify that reliable end-to-end automatic localization and segmentation within complete CBCT examinations were not achieved, and describe the VOI-based approach as preliminary computer-assisted segmentation within preselected tooth volumes. We also added the poor full-volume performance and the absence of further model optimization as limitations. Finally, we expanded the discussion of future improvement strategies, including two-stage localization–segmentation frameworks, inclusion of pulp-stone-negative tooth VOIs, increased representation of small calcifications, small-object-sensitive sampling and loss functions, multicenter datasets, and external validation.
Page 9, lines 324–336; pages 14–16, lines 420–524; pages 17–18, lines 561–612; pages 18–19, lines 619–628.
Comment 4- Please provide high-res figures with increased text font size for better visualization.
Response: We thank the reviewer for this helpful recommendation. In response to this comment, the figures were revised to improve their clarity and suitability for publication. Image resolution was increased, and the font sizes of titles, axis labels, tick labels, legends, annotations, and flowchart text were enlarged. Figure dimensions, spacing, and element positioning were also adjusted to prevent overlap among labels, numerical values, arrows, axes, and plotted elements. The revised high-resolution figures have been incorporated into the manuscript.
Revised Figures 1–5: Figure 1, pages 4–5, lines 168–172; Figure 2, page 11, lines 382–383; Figure 3, page 12, lines 392–393; Figure 4, page 13, lines 402–406; Figure 5, pages 13–14, lines 408–418.
Comment 5- According to the ROC curve in Figure 4, the trained model has very high specificity and low sensitivity. Please provide a scientific reason for this observation.
Response: We appreciate the reviewer’s observation regarding the marked difference between specificity and sensitivity. The very high voxel-level specificity and comparatively lower sensitivity can most plausibly be explained by severe voxel-level class imbalance and the model’s relatively conservative foreground predictions. Background voxels substantially outnumbered pulp-stone voxels within the tooth-level VOIs. Consequently, most background voxels were correctly classified and relatively few false-positive voxels were generated, resulting in a specificity of 0.999. In contrast, pulp stones were small and, in some cases, represented by only a few voxels because the native voxel size ranged from 0.3 to 0.4 mm. Partial-volume effects, CBCT noise, indistinct lesion boundaries, and radiodensity similarities between pulp stones and adjacent dentin may have increased false-negative predictions, resulting in a sensitivity of 0.635.
This explanation has been added to the Discussion. We have also clarified in the Methods, Results, Figure 4 legend, and Limitations that sensitivity and specificity were calculated at the voxel level within preselected pulp-stone-positive tooth VOIs. Consequently, the reported specificity reflects the correct classification of background voxels within positive VOIs and should not be interpreted as tooth-level diagnostic specificity for identifying pulp-stone-negative teeth.
Pages 7–8, lines 274–302; page 9, lines 338–346; page 13, lines 402–406; page 15, lines 449–458; page 18, lines 593–599.
Comment 6- Please provide images of the predictions alongside the ground-truth for comparison with different test datasets.
Response: We appreciate the reviewer’s helpful suggestion. Because the study included a single held-out VOI-level test dataset, visual comparisons were performed using representative cases with high, moderate, and low segmentation performance from this dataset. Accordingly, Figure 5 has been revised to present side-by-side comparisons of the original cropped CBCT images, manually generated ground-truth segmentations, and corresponding model predictions for these cases. The revised figure enables direct visual assessment of the agreement and discrepancies between the reference annotations and model-generated segmentations across different performance levels. The related text in the Results section and the Figure 5 legend have also been updated accordingly.
Page 13, lines 407–412; pages 13–14, lines 408–418.
Reviewer 2 Report
Comments and Suggestions for AuthorsIn line 61, the authors stated: "Periapical and bitewing radiographs have traditionally been used for the radiographic assessment of pulp stones", which is absolutely correct. However, later they use a panoramic X-ray to perform an early check for the presence or absence of pulp stones. This is an inconsistent approach and needs some explanation. I believe a donwsize of that approach should also be noted, as for some patients, including those with skeletal Class II and Class III anterior teeth, may appear out of focus. In such circumstances, panoramic image early screening may not always be the best choice.
I would rather avoid the term "routine" in radiology (line 126), as we all try not to perform routine examinations but rather standardised ones based on a patient's individual needs.
Little is known about whether the model works better with anterior or posterior teeth, in the maxilla or the mandible.
In the limitations section, one should also include the study's design, specifically the use of only larger FOVs in CBCT examinations. Smaller FOVs result in smaller pixel sizes, which may be influential for the perfomrance of the proposed approach, perhaps with no need to generate tooth-lever ROIs.
For clinicians who perform endodontic treatment, the smaller FOV will also be a first choice because of the image quality.
Author Response
We sincerely appreciate the reviewer’s thoughtful and clinically relevant comments. These observations helped us clarify the radiographic selection procedure, use more appropriate radiological terminology, report the anatomical composition of the test set, and more explicitly acknowledge the limitations associated with the CBCT acquisition protocol. The manuscript was revised as detailed below.
Comment 1: The use of panoramic radiography for initial screening appears inconsistent with the statement that periapical and bitewing radiographs have traditionally been used for assessing pulp stones. The limitations of panoramic screening, particularly the potential blurring of anterior teeth in patients with skeletal Class II or Class III relationships, should be explained.
Response: We appreciate the opportunity to clarify the role of panoramic radiography in the study. Panoramic radiographs were used as an initial screening tool because paired panoramic radiographs and CBCT examinations were available in the retrospective institutional archive. Panoramic screening was not considered the definitive reference standard; pulp-stone presence and three-dimensional localization were subsequently evaluated on multiplanar CBCT images by the observer panel. This rationale has been clarified in the Methods.
We also acknowledge the limitations of this selection procedure. The Limitations section now states that panoramic screening may have excluded small or subtle pulp stones visible only on CBCT and that anterior teeth located outside the focal trough because of patient positioning or skeletal relationships may appear blurred. Moreover, requiring visibility on both modalities may have preferentially selected larger or more radiopaque calcifications and introduced selection bias. The Discussion additionally emphasizes that panoramic radiography should not be regarded as the preferred modality for definitive three-dimensional assessment.
Page 4, lines 143–151; page 5, lines 175–180; page 16, lines 485–492; page 17, lines 571–575.
Comment 2: The term “routine” should be avoided because radiographic examinations should be standardized and performed according to individual patient needs.
Response: This terminological concern is well founded. The term “routine” has been removed from the manuscript. The revised Methods section now states that the CBCT examinations were obtained using standardized acquisition parameters selected according to the clinical indication and individual diagnostic needs. This wording more accurately reflects the principle that radiographic examinations must be clinically justified and tailored to each patient.
Pages 3–4, lines 133–139.
Comment 3: It remains unclear whether the model performs differently in anterior and posterior teeth or in the maxilla and mandible.
Response: The reviewer raises an important question concerning anatomical variation in model performance. We have therefore added the anatomical distribution of the held-out VOI-level test set to the Results. Of the 23 test VOIs, 15 were obtained from maxillary teeth and 8 from mandibular teeth, while 1 represented an anterior tooth and 22 represented posterior teeth. Because this distribution was markedly unbalanced, particularly because only one anterior tooth was included, reliable subgroup comparisons according to dental arch or tooth location could not be performed. This issue is now explicitly acknowledged in the Results and Limitations, and future studies with sufficiently balanced anatomical subgroups are recommended.
Page 10, lines 353–361; page 18, lines 589–612.
Comment 4: The use of only large-FOV CBCT examinations should be acknowledged as a limitation. Limited-FOV examinations may permit smaller voxel sizes, potentially improve model performance, and reduce the need for tooth-level VOIs. Limited FOV is also generally preferred when CBCT is indicated for endodontic assessment.
Response: We agree that the acquisition protocol is important when interpreting and generalizing the present findings. The Methods section now specifies that all CBCT examinations were acquired using a 16 × 13 cm FOV with native voxel sizes ranging from 0.3 to 0.4 mm. The Limitations section explains that these parameters may have limited the visibility and representation of small pulp stones because some calcifications may occupy only a few voxels and may therefore be particularly susceptible to partial-volume effects. We have also clarified that cropping tooth-level VOIs reduces irrelevant anatomical backgrounds but does not improve the intrinsic spatial resolution of the original CBCT data.
The Discussion now notes that, when CBCT is clinically indicated for endodontic assessment, a limited FOV appropriate to the region under investigation is generally preferred to restrict the irradiated volume and may permit smaller voxel sizes. Accordingly, the findings may not be directly generalizable to limited-FOV endodontic protocols. Future studies should compare full-volume and tooth-level VOI approaches across large- and limited-FOV examinations acquired with different native voxel sizes to determine whether higher-resolution imaging improves small-calcification segmentation and reduces dependence on prior tooth-level localization.
Pages 3–4, lines 135–139; page 16, lines 485–501; page 18, lines 578–582 and 605–612.
Reviewer 3 Report
Comments and Suggestions for AuthorsThis is an interesting study addressing the automatic three-dimensional segmentation of pulp stones on CBCT images using an nnU-Net v2 model. The topic is relevant and the methodology is generally well described.
However, there are several important issues that should be addressed before the manuscript can be considered for publication. Most importantly, the actual experimental design does not fully support some of the claims regarding automatic detection and diagnostic performance.
Authors should carefully read the following comments and modify the manuscript accordingly.
1) The main limitation is that the ROI-based model does not actually evaluate automatic detection of pulp stones. The final analysis was performed on tooth ROIs obtained exclusively from teeth already known to contain pulp stones. Therefore, the model evaluates the ability to segment a pulp stone once the affected tooth has already been identified, rather than the ability to automatically detect pulp stones in an unselected CBCT examination. This distinction is important and should be clearly stated throughout the manuscript. In particular, references to “automatic detection”, “diagnostic accuracy”, and clinical decision support should be reconsidered.
The Abstract, Discussion and Conclusions should be modified accordingly, and the authors should consider whether the term “Diagnostic Accuracy” is appropriate in the title.
2) Another main problem is that the ROI dataset apparently contains no true negative teeth. The very high specificity and overall accuracy therefore represent voxel-level classification of pulp-stone versus background voxels within positive tooth ROIs. They should not be interpreted as the ability of the model to distinguish teeth with pulp stones from teeth without pulp stones.
The authors should clearly explain this limitation and modify the interpretation of specificity and accuracy accordingly. Without evaluation on negative teeth, clinical diagnostic accuracy cannot be established.
3) Authors emphasized on accuracy and AUC, however considering the class imbalance and the remaining segmentation errors this emphasis is excessive. The model achieved an accuracy of 0.998 and AUC of 0.967, but Dice was 0.637, precision 0.695 and recall 0.635. Thus, segmentation performance is clearly more moderate than the accuracy and AUC values might suggest.
The authors should place less emphasis on accuracy and ROC/AUC and give greater weight to Dice, precision, recall and surface-distance measures. Statements describing the model as having “excellent discriminative performance” should be moderated.
4) Only 23 tooth-level ROIs were included, and Dice values ranged from 0 to 0.838.
The authors should indicate how many individual patients contributed to these 23 ROIs and provide confidence intervals for the main performance estimates. The limited size of the test sample should also be more clearly acknowledged when interpreting the results.
5) The authors state that training was performed using the fold=all configuration, with all training data being used, but subsequently state that the model with the “best validation performance” was selected. The training and validation procedure is unclear.
Please, authors should clarify exactly how internal validation and model selection were performed and clearly distinguish the training, validation and independent test datasets.
6) The full-volume model showed poor performance (DSC 0.183, precision 0.254 and recall 0.205), and the supplementary analysis reports an AUC of only 0.60. Performance improved substantially only after restricting the analysis to tooth ROIs already known to contain pulp stones. This finding is highly relevant because the full-volume approach more closely resembles the intended clinical application.
The authors should discuss this question more explicitly and acknowledge that automatic localization/detection within complete CBCT volumes remains unresolved.
7) The authors should also clarify why pulp stones were required to be visible on both panoramic radiographs and CBCT. This criterion may preferentially select larger or more radiopaque pulp stones and should be justified and discussed as a potential source of selection bias.
8) If available, interobserver agreement before consensus should also be reported, as consensus segmentation alone does not provide information on the reproducibility of the reference standard.
Author Response
We sincerely appreciate the reviewer’s careful and constructive evaluation of our study. The comments drew our attention to important distinctions between voxel-level segmentation, automatic tooth-level detection, and clinical diagnostic performance. In response, we substantially revised the title, Abstract, Methods, Results, Discussion, Limitations, Conclusions, tables, and figure legends to ensure that the scope and interpretation of the findings accurately reflect the experimental design.
Comment 1
The main limitation is that the ROI-based model does not actually evaluate automatic detection of pulp stones. The final analysis was performed on tooth ROIs obtained exclusively from teeth already known to contain pulp stones. Therefore, the model evaluates the ability to segment a pulp stone once the affected tooth has already been identified, rather than the ability to automatically detect pulp stones in an unselected CBCT examination. This distinction is important and should be clearly stated throughout the manuscript. In particular, references to “automatic detection,” “diagnostic accuracy,” and clinical decision support should be reconsidered. The Abstract, Discussion and Conclusions should be modified accordingly, and the authors should consider whether the term “Diagnostic Accuracy” is appropriate in the title.
Response: This important distinction has been addressed throughout the revised manuscript. The title was changed from one emphasizing “Diagnostic Accuracy” and “Automatic Segmentation” to
“Three-Dimensional Segmentation of Pulp Stones in Tooth-Level Volumes on Cone-Beam Computed Tomography Using nnU-Net v2: A Preliminary Study.”
The Abstract, Introduction, Methods, Discussion, and Conclusions were revised to clarify that the primary VOI-based analysis evaluated voxel-wise segmentation within preselected tooth-level volumes already known to contain pulp stones. The manuscript now explicitly states that this analysis did not evaluate automatic tooth-level detection in unselected CBCT examinations. References suggesting established diagnostic accuracy, end-to-end automatic detection, clinical readiness, or validated clinical decision support were removed or moderated. The proposed approach is now described as preliminary computer-assisted pulp-stone delineation within preselected tooth volumes. We also emphasize that reliable automatic localization within complete CBCT examinations remains unresolved.
Page 1, lines 2–43; page 3, lines 95–108; pages 14–15, lines 419–467; pages 18–19, lines 619–628.
Comment 2
Another main problem is that the ROI dataset apparently contains no true negative teeth. The very high specificity and overall accuracy therefore represent voxel-level classification of pulp-stone versus background voxels within positive tooth ROIs. They should not be interpreted as the ability of the model to distinguish teeth with pulp stones from teeth without pulp stones. The authors should clearly explain this limitation and modify the interpretation of specificity and accuracy accordingly. Without evaluation on negative teeth, clinical diagnostic accuracy cannot be established.
Response: We agree that the original interpretation required clarification. The terminology has been corrected from ROI to VOI, and the Methods now explicitly state that all 238 tooth-level VOIs were obtained from teeth already known to contain at least one pulp stone and that no pulp-stone-negative tooth VOIs were included.
The Methods, Results, Discussion, Limitations, table footnotes, and Figure 4 legend were revised to clarify that sensitivity, specificity, accuracy, and ROC–AUC were calculated at the voxel level within preselected pulp-stone-positive tooth VOIs. In this setting, true-negative voxels represent correctly classified background voxels within positive tooth VOIs rather than pulp-stone-negative teeth. Accordingly, these metrics cannot be interpreted as tooth-level diagnostic sensitivity, specificity, accuracy, or discrimination. The revised manuscript explicitly acknowledges that the model’s ability to distinguish pulp-stone-positive from pulp-stone-negative teeth and its clinical diagnostic accuracy remain unknown.
Page 6, lines 222–230; pages 7–8, lines 274–302; pages 9–10, lines 338–374; page 13, lines 394–406; pages 14–15, lines 420–458; page 18, lines 593–599.
Comment 3
Authors emphasized accuracy and AUC; however, considering the class imbalance and the remaining segmentation errors, this emphasis is excessive. The model achieved an accuracy of 0.998 and AUC of 0.967, but Dice was 0.637, precision 0.695, and recall 0.635. Thus, segmentation performance is clearly more moderate than the accuracy and AUC values might suggest. The authors should place less emphasis on accuracy and ROC/AUC and give greater weight to Dice, precision, recall, and surface-distance measures. Statements describing the model as having “excellent discriminative performance” should be moderated.
Response: The reviewer’s concern regarding the influence of class imbalance is well founded. We removed statements characterizing the model’s discriminative performance as excellent and reduced the emphasis placed on voxel-level accuracy and ROC–AUC. The revised Results and Discussion give greater weight to the Dice similarity coefficient, Jaccard index, precision, recall, false-positive and false-negative predictions, and surface-distance measurements.
The VOI-based performance is now consistently characterized as moderate. We also explain that the high voxel-level accuracy was strongly influenced by the predominance of background voxels and that ROC–AUC reflects pulp-stone-versus-background voxel discrimination within positive VOIs rather than tooth-level diagnostic discrimination. The revised interpretation emphasizes that overlap, error-based, classification, and surface-distance metrics should be considered together rather than relying on accuracy or AUC alone.
Page 1, lines 28–43; pages 9–10, lines 324–374; pages 12–13, lines 394–406; pages 14–16, lines 420–518.
Comment 4
Only 23 tooth-level ROIs were included, and Dice values ranged from 0 to 0.838. The authors should indicate how many individual patients contributed to these 23 ROIs and provide confidence intervals for the main performance estimates. The limited size of the test sample should also be more clearly acknowledged when interpreting the results.
Response: Further information has been added to clarify the composition and statistical uncertainty of the test set. The revised Results state that the 23 tooth-level VOIs were obtained from 19 individual patients and that case-level Dice values ranged from 0.000 to 0.838. The mean Dice similarity coefficient is now reported with its 95% confidence interval as 0.637 (95% CI: 0.560–0.693).
The confidence interval was calculated using patient-level cluster bootstrap resampling with 10,000 repetitions to account for the inclusion of more than one test VOI from some patients. The resampling procedure has been described in the Methods. Because case-level values were available only for the Dice similarity coefficient, a confidence interval could be calculated only for this principal overlap measure. The small test sample is now explicitly acknowledged in the Limitations as restricting the precision and generalizability of the performance estimates.
We have also clarified that allocation of the cropped dataset was performed at the tooth-VOI level rather than the patient level. Although none of the 23 test VOIs was used for model fitting, other VOIs from some of the same patients could have been included in the training dataset. This potential patient-level overlap is now reported in the Abstract, Methods, Results, Discussion, Limitations, and Figure 1 legend and is acknowledged as a possible source of optimistic performance estimates.
Page 1, lines 22–25; pages 6–8, lines 231–238 and 309–313; pages 10–11, lines 353–383; page 17, lines 561–570.
Comment 5
The authors state that training was performed using the fold=all configuration, with all training data being used, but subsequently state that the model with the “best validation performance” was selected. The training and validation procedure is unclear. The authors should clarify exactly how internal validation and model selection were performed and clearly distinguish the training, validation and independent test datasets.
Response: We appreciate the opportunity to correct this ambiguity. The previous statement indicating that the model with the “best validation performance” was selected has been removed because it incorrectly implied the use of a separate validation dataset.
The revised Methods clarify that the 238 cropped tooth-level VOIs were divided at the VOI level into 215 VOIs for model training and 23 VOIs for held-out testing. The final nnU-Net v2 model was trained for 1,000 epochs using the fold=all configuration, in which all 215 training VOIs were used for model fitting. No separate internal validation subset was created, and no validation-based model selection was performed. After completion of training, the final trained model was evaluated using the 23 held-out VOIs, none of which had been used for model fitting.
Because allocation was performed at the VOI level, patient-level independence between the training and test datasets was not ensured. Therefore, the term “independent test set” has been replaced with “held-out VOI-level test set” when referring to the VOI-based analysis. The full-volume analysis, for which CBCT examinations were allocated exclusively at the patient level, is described separately.
Pages 6–7, lines 218–238 and 265–272.
Comment 6
The full-volume model showed poor performance (DSC 0.183, precision 0.254 and recall 0.205), and the supplementary analysis reports an AUC of only 0.60. Performance improved substantially only after restricting the analysis to tooth ROIs already known to contain pulp stones. This finding is highly relevant because the full-volume approach more closely resembles the intended clinical application. The authors should discuss this question more explicitly and acknowledge that automatic localization/detection within complete CBCT volumes remains unresolved.
Response: This clinically important observation is now emphasized more explicitly. The Results report the poor full-volume Dice similarity coefficient of 0.183, precision of 0.254, recall of 0.205, and supplementary voxel-level AUC of 0.60. The revised Discussion acknowledges that the full-volume setting more closely represents the intended clinical scenario because the location of the affected tooth is not known in advance.
We now state clearly that reliable automatic localization and segmentation of pulp stones within complete CBCT examinations were not achieved. Although restricting the input to preselected tooth-level VOIs reduced irrelevant background and improved performance, it required prior identification and localization of the pulp-stone-positive tooth. Consequently, the VOI-based approach is not presented as an end-to-end detection system. This limitation is also emphasized in the Limitations and Conclusions, and the model is described as preliminary and unsuitable for independent clinical use in its current form.
Page 1, lines 28–43; page 9, lines 324–336; pages 14–15, lines 420–467; page 18, lines 582–588; pages 18–19, lines 619–628.
Comment 7
The authors should also clarify why pulp stones were required to be visible on both panoramic radiographs and CBCT. This criterion may preferentially select larger or more radiopaque pulp stones and should be justified and discussed as a potential source of selection bias.
Response: The rationale for this eligibility criterion has been clarified in the Methods. Visibility on both modalities was used as a conservative case-selection criterion to ensure that the calcifications could be confidently identified and matched across paired panoramic and CBCT examinations before three-dimensional annotation. Panoramic radiography was not considered the definitive reference standard; three-dimensional assessment and annotation were performed using multiplanar CBCT images and consensus review.
We acknowledge that this criterion may have excluded small or subtle pulp stones visible only on CBCT and may have preferentially selected larger or more radiopaque calcifications. The resulting risk of selection bias is now explicitly discussed in the Limitations.
Page 4, lines 143–151; page 5, lines 175–180; page 16, lines 485–492; page 17, lines 571–575.
Comment 8
If available, interobserver agreement before consensus should also be reported, as consensus segmentation alone does not provide information on the reproducibility of the reference standard.
Response: The reviewer raises an important point regarding the reproducibility of the reference standard. In this study, the observers jointly reviewed pulp-stone presence and segmentation boundaries, and disagreements were resolved through consensus. Separate observer-specific voxel-wise segmentation masks were not generated before consensus. Consequently, quantitative interobserver agreement could not be calculated retrospectively.
The Methods have been revised to describe the consensus procedure more clearly, and the inability to assess quantitative interobserver agreement has been explicitly acknowledged as a limitation. We agree that future studies should obtain separate observer-specific annotations before consensus to permit formal assessment of reference-standard reproducibility.
Pages 5–6, lines 185–206; page 17, lines 575–577.
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsDear Authors,
the revised manuscript was improved significantly and in this form is OK for Me.
Best regards.
Reviewer 3 Report
Comments and Suggestions for AuthorsThe authors have responded appropriately to the reviewers' comments, and the corresponding changes have been satisfactorily incorporated into the revised manuscript. I have no further major comments, and I consider the manuscript suitable for publication in its present form.