Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Authors presented the accuracy and reproducibility of digital 2D and 3D nuclear segmentation algorithms, and claimed that the 3D approache can overcome thresholding and 2D constraints. They detected the performance of the adaptive threshold-based FISHQuant and the AI-powered Cellpose, NucleAIzer and StarDist algorithms, which highlighting the advantage of deep learning-based methods over traditional thresholding or manual counting for safe nuclear segmentation. These comparative results will offer implications for FISH-based research. I just have several minor concerns as followed:

1. In the results section, all of Table title are too long, some detailed information should be edited as “note or annotation”;
2. The discussion section could be further disscussed additional aspects of applying 3D algorithms in FISH research. Such as, expand on how deep-learning techniques enhance the role of  automating signal analysis, and discussed challenges like handling noise in  adapting algorithms for different tissue types.

Author Response

Thank you for your recommendations any criticisms!

You can find our modification to the original manuscrip in red color.

As you rightly suggested we cut the titles not to be comprehensively analyzing the results of all table
legends.

We added a paragraph to the starts of the our discussion and another at its end on upstrem techical aspects
of digitalization and the tested algorithms as suggested, respectively.

Although several deep learning algorithms for spot detection are available (Bouilhol 2022;Xue 2023;Jian
2024), their integration with nuclear segmentation frameworks remains essential, particularly because
only the signals localized within nuclei are diagnostically relevant (Wolff, 2007). The present work
focused primarily on nuclear detection and its clinical evaluation, as accurate identification of nuclei is
the prerequisite for reliable downstream FISH analysis. While deep neural networks provide multiple
options for spot detection, no unified framework is currently accessible that allows simultaneous
execution of nuclear and spot detection algorithms in an out of the box manner. To address this limitation,
a custom spot detection application has been under development, designed to identify signals specifically
within the masks of segmented nuclei across the spot channels (Kuczarov 2024). Our intention here was to
test contemportary and freely accessible deep-learning algorithms if they can improve the nuclear
segmentation accuracy of our adaptive thresholding FISHQuant algorithm of diagnostic potential
(Kiszler, 2013; Chea, 2017).

Fluorescence microscopy may suffer from photon shot noise, camera readout artifacts, and
autofluorescence, which impose strict constraints on spatial fidelity and quantitative accuracy. Supervised
deep learning methods such as CNNs, U-Nets, and transformer hybrids utilize paired training data to
learn strong priors, though they often face challenges in generalization across experiments (Ronneberger
et al., 2015; Moen et al., 2019). Self-supervised strategies including Noise2Noise, Noise2Void, and blind-
spot networks reduce data requirements and improve adaptability by exploiting redundancy without clean
targets (Lehtinen et al., 2018; Krull et al., 2019; Batson & Royer, 2019). Modality-specific AI corrections
and hybrid wavelet–deep learning pipelines may further enhance denoising performance, enabling
improved SNR while preserving fine biological structures (Zhou et al., 2024; Müller et al., 2025).

A new, Table 5 was inserted.

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript by Csizmadia and co-authors “Enhanced Precision of Fluorescent In Situ Hybridization (FISH) Analysis Using Neural Network-Based Nuclear Segmentation For Digital Microscopy Samples” addresses a significant challenge in digital pathology: the accurate segmentation of overlapping nuclei in dense tissue samples for FISH analysis. The comparison of traditional thresholding methods with state-of-the-art deep learning algorithms (Cellpose, NucleAIzer, StarDist) in both 2D and 3D contexts is a valuable contribution. The manuscript is generally well-structured, and the research question is clearly defined. However, the study has several points to be addressed before it can be accepted for publication.

Major points.

1. The choice of  follicular lymphoma, known for its high nuclear density, is optimal for demonstration of limitations of 2D segmentation algorithms. However, the study is based only on 5 follicular lymphoma cases, which limits the statistical power of the findings and rises concerns of overfitting. So, the authors should explicitly state these limitations in the discussion or, better yet, validate the findings on an independent larger dataset.
2. All the algorithms are compared to the manual evaluation, which is considered a “gold standard”. However, this method systematically gives lower nuclei counts, meaning underestimation of number of nuclei in crowded regions. This suggests that the manual method is likely least accurate and cannot be used as a benchmark due to potential misrepresentation  of the performance of automated methods. In this case, the best approach seems to be a manual annotation of 3D images, which may not be feasible but should be acknowledged as a limitation. Anyway, instead of treating manual counting as an absolute truth, the authors should consider is a clinical practice baseline.
3. Cellpose is introduced and used in initial 2D segmentation but is then dropped from the critical 3D and genetic classification analyses without a clear, data-driven justification in the results section. The justification appears later in the discussion. This selective reporting weakens the comparative analysis. Its exclusion should be justified upfront in the results with the supporting data (e.g., "Cellpose was excluded from further 3D analysis due to its consistently low performance in initial 2D tests and failure to detect low-intensity signals...").
4. The interpretation of the Pearson correlation matrix (Table 1) is confusing. The text states "Strong inverse correlations were observed between NucleAIzer and manual evaluation," which is correct (r = -0.839), but then it's unclear what a "negative correlation" means in the context of nucleus counting. It simply means when one method counts high, the other counts low, reinforcing that manual counting is an outlier.
5. The variance values for StarDist (VS ~13,000-14,000) are an order of magnitude higher than for NucleAIzer and FISHQuant (VS ~1,200-1,700). The authors note this but do not sufficiently discuss the potential causes or implications. Such high variance suggests StarDist is highly unstable and its results may not be reliable, which contradicts its high F-measure and Accuracy scores. This discrepancy needs to be resolved and explained.
6. The statement "The only 4% differences in VP/VS ratio" is unclear. A ratio close to 1 is expected (since VP ≈ VS for large n), and it does not necessarily indicate "consistent and comparable performance," especially when the absolute variance values are vastly different.
7. While the study uses Z-stacks, the analysis of 3D segmentation is primarily presented as nucleus counts per plane. There is no quantitative evaluation of the quality of the 3D segmentation itself (e.g., using 3D Dice coefficients or surface distance metrics against a manually segmented 3D ground truth). Figure 5C is a visual example but lacks quantitative support.A quantitative 3D segmentation accuracy assessment would significantly strengthen the claim that 3D AI methods are superior.

 

Minor ponts.

• All microscopic images lack the scale bars.
• Line 93  - "The Cellpose program... pre-trained on over 70,000 diverse fluorescent cellular objects, enables 3D nuclear segmentation by utilizing two-dimensional (2D) model data." This is a confusing statement. It should be clarified that it uses a 2D model applied to 3D data by processing orthogonal slices.
• The choice of the plot style in Figure 4, B is weird for nominal data, please consider bar charts.

 

Comments on the Quality of English Language

• The manuscript should be thoroughly proofread for numerous typos:

1. line 20-  "Fluorescent In Situ Hybridization" is the standard term, but the abstract has a typo: "fluorescence in suit hybridization".
2. line 33  "Pathogenetic classification" should likely be "pathogenic classification" or "pathogenetic alteration classification".
3. line 43 - "precisely analyses" -> "precisely analyzed".
4. line 49 - "improving the accuracy automated digital separation" -> "improving the accuracy of automated digital separation".
5. Throughout the text: “Startdist" / "Statdist" -> "StarDist" (be consistent with the correct capitalization).  
6. Line 147 - "a mountant in countaining DAPI" -> "in a mounting medium containing DAPI".
7. Line 309 -  "details cell nuclei separation" -> "details  of cell nuclei separation". Line  468 -  "fall short in Accuracy to" -> "fell short in Accuracy  compared to". 
8. Line 445 - “segmentad” - > “segmented”
9. Line 492 - "The latter approaches, particularly the NucleAIzer and Cellpose algorithms can be offered..." Given that Cellpose performed poorly, this sentence is misleading. It should focus on NucleAIzer and StarDist.   
10. Line 528 - "achieving high F-measure (1.01)" - An F1-score cannot be greater than 1.0. This is a critical error that must be checked and corrected in the data of Table 4.

 

Author Response

Thank you for your detailed views and suggestions to improve our manuscript!

Your can find our modifications to the original manuscrip in red color.

Major points.
1. We agree that 5 cases tested with 2 different methods (10 samples) are low numbers, which we
mention as a limitation in the revised manuscipt. However, the detected 500-1000 nuclei per
sample depending on the method used on average 4000 nuclei per each method shold be enough

for a valid statistical comparison. Nevertheles, in the revised discussion we mentioned the case
number as a potential limitation of our study.

2. The eye-control based method was choosen as a gold standard, since most diagnostic laboratories
still use traditional analysis of at least 50-200 tumor cells per sample depending on the type of
tumor and gene aber-ration required by the guidelines for a reliable diagnostic decision making
(Wolff, 2007; Mascarello 2011). This fact and the frequent overlapping nuclei in lymphoma
sections emphasized the need to search for au-tomation and compare the analytic performane of
FISHQuant and the AI-based, general purpose CellPose, NucleAIzer and Stardist algorihms in
order to find that, which can improve the accuracy and reproduciblitity of FISH reading and
reduce the chance for bias and fatigue of pathologists.
We agree to state early that "Cellpose was excluded from further 3D analysis due to its consistently low
performance in initial 2D tests and failure to detect low-intensity signals".
In the context of nucleus counting, a negative correlation means that when manual evaluation identifies
fewer nuclei than NucleAIzer. In other words, the two methods move in opposite directions: a decrease in
nuclei detected by one method corresponds to an increase in the other.

The variance values …
In our comparative analysis of nucleus segmentation algorithms across multiple runs and sample
conditions, StarDist consistently reported the highest number of detected nuclei. For example, in sample
BCL2_2661_2, StarDist detected 942, 845, 932, 943 and 942 nuclei across five runs; whereas FISHQuant
reported 801, 777, 788, 801, and 800, and NucleAIzer reported 735, 734, 733, 736, and 737, while eye-
controlled analysis safely found only ~500 . This pattern was observed across all samples: StarDist
produced counts in the range of 850–1000 nuclei, while FISHQuant and NucleAIzer remained closer to
700–800 nuclei.
The consistently higher baseline counts explain why StarDist exhibits larger population variance (VP) and
sample variance (VS). Variance is calculated from squared deviations from the mean, and when the mean
values are high, even small fluctuations translate into larger absolute deviations. For instance, a ±40
change around 950 nuclei contributes more strongly to variance than a ±20 change around 730 nuclei.
Thus, StarDist’s elevated counts naturally amplify both VP and VS, even though its relative stability
across runs appears similar to the other algorithms.
Population variance (VP) reflects the spread of values across the entire dataset. Because StarDist
consistently detects more nuclei than the other methods, its values sit further from the overall mean,
inflating VP. Sample variance (VS), which measures variability within repeated runs of the same sample,
is also higher for StarDist. Although the counts remain stable in relative terms, the absolute deviations
are larger due to the higher baseline. This explains why StarDist shows higher VS despite maintaining
consistent detection across runs.
In summary, StarDist’s higher population and sample variances are not indicators of instability but rather
a mathematical consequence of its consistently elevated nucleus counts. The algorithm’s segmentation
strategy leads to robust detection of nuclear objects, but the scale of its output magnifies variance
measures compared to FISHQuant and NucleAIzer. These findings highlight the importance of
interpreting variance in the context of baseline counts: StarDist’s higher VP and VS reflect its detection
magnitude rather than inconsistency in performance.
The only 4% differences …

Our calcualtion formula was:
The phrase “4% difference” refers to that VP values are about 4% lower than the corresponding VS
values, resulting in a ratio of ~96%, which would be ~1 in relative trend.

While the study uses Z-stacks…,
We acknowledge that metrics such as Dice coefficients or surface distance measures would provide
additional validation. However, since the segmentation algorithms (Cellpose, StarDist, NucleAIzer) were
executed in different environments—ImageJ, QuPath, and server‑side batch processing—it was not
feasible to apply a unified 3D statistical procedure. Each platform produced outputs with distinct formats,
scaling, and post‑processing rules, making pooled quantitative analysis unreliable. Instead, we
emphasized the number of identified nuclei as a stable metric and used this directly in the FISH
evaluation, where accurate signal interpretation depends on consistent nucleus detection. This approach
avoided bias from incompatible 3D frameworks and ensured biologically meaningful results, in line with
established practices that prioritize nucleus counts over volumetric metrics when environments differ
(Graham et al., 2021; Schmidt et al., 2018).

Graham, S., et al. Evaluation of nuclei segmentation in computational pathology: methods and
challenges. Medical Image Analysis, 2021.
Schmidt, U., et al. Cell segmentation: a benchmark and evaluation framework. Nature Methods, 2018.

We also agree, that quantitative accuracy metrics would further strengthen this claim, our study focused
on demonstrating the diagnostic relevance of 3D segmentation within the FISH classification workflow,
which we believe provides meaningful evidence of its added value.

Minor ponts.
Figures were added scale bars.
We state in the revised manuscriopt that “All free-access general purpose AI tools were used off-the-shelf,
without any further training method.”
Yes, Fig 4A-B may be unusual but we think that they can demontrate the relation amongst different
methods.
All typos including those mentioned are corrected in the revised manuscript.

A new, Table 5. was inserted.

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript by Csizmadia et al., evaluates and compares AI-based 3D nuclear segmentation (NucleAIzer, StarDist, Cellpose) against 2D adaptive thresholding (FISHQuant) on multilayer Z-stacks of FL FISH (BCL2). Authors finds that 3D deep-learning methods—especially NucleAIzer/StarDist—detect more nuclei with lower run-to-run variance and stronger gene-spot correlations than 2D/manual approaches.

However, the claim of improved 3D segmentation with deep learning is not novel per se; prior literature and even the manuscript’s own citations already establish robust DL nuclei segmentation (e.g., NucleAIzer, StarDist, Cellpose families). The added value here could be the FL FISH use-case with BCL2 probes and multilayer Z-stacks, but with n=5, this remains preliminary. Moreover, at present, this work requires major revision before reconsideration.

Major issues to address:

1. Authors present too many broad claims, and missed on specifying a focused hypothesis and a single primary endpoint (e.g., accuracy vs. F-measure vs. VP/VS) in the abstract.

2. The manuscript fails to define what “diagnostic reliability” means in a measurable way in the abstract; statements like “improved gene spot correlation (r>0.82)” are given without specifying reference standard or dataset size.

1. Introduction should include a concise rationale for selecting NucleAIzer/StarDist/Cellpose and for choosing 3D over 2D for this clinical task, tied to a pre-stated aim.

4. Authors should have described patient/sample selection with more detail (n=5 FL cases only; inclusion criteria, case heterogeneity, signal quality strata), and this small sample size is a major limitation for generalization.

 

5. Authors did not provide full parameter settings, preprocessing, or training/adaptation details for StarDist/NucleAIzer/Cellpose (thresholds, model variants, augmentation, post-processing), making reproducibility weak. Versions are partly listed, but not the exact configs.

1. It is unclear whether Cellpose was re-trained on domain data or used off-the-shelf, affecting fairness of comparisons.

7. Similarly, authors should have clarified who performed manual annotations, how many readers, blinding, and inter-rater metrics; “eye-controlled” counting limited to “obviously separated nuclei” is an inadequate gold standard and biases results.

 

8. Primary comparison and statistical plan are not pre-specified; correlations, VP/VS, and classification metrics are mixed without hierarchy, and negative correlations in Table 1 are confusing.

1. Please define sample vs. population variance computation precisely (formulas, units) yet relies on VP/VS ≈ 0.96 to claim stability; this needs rigorous statistical description.
2. Manuscript fails to separate the results for detection (nuclei counts) from classification (spot patterns) with clear ground-truthing; using a gold standard that ignores overlapping nuclei while AI does not will naturally inflate AI counts and is not proof of correctness.
3. Authors should have tempered claims in the Discussion about clinical integration and “diagnostic reliability,” given the tiny dataset (five cases), incomplete gold standard, and limited external validation.
4. Authors did not separate analytic validity (segmentation accuracy) from clinical validity (correct genetic classification vs. an independent expert consensus) and clinical utility (impact on decisions).
5. The manuscript hasn’t acknowledged possible circularity: the same pipeline that defines nuclei then measures spots within those masks, which can inflate apparent internal consistency.
6. Authors should also acknowledge that upstream imaging factors (fluorophore choice, autofluorescence, clearing chemistry, detector spectra) drive S/N and can bias segmentation and spot-calling; the study by Kumar, V. et al. 2021, Brain Struct Funct 226, 481–499, shows far-red dyes (e.g., AF-647) consistently yield higher S/N than AF-488/594, with weaker channels further degraded by clearing solvents and instrumentation.
7. Correct administrative details; the ethics approval date appears malformed (“issued in 10/10/205”).

Minor issues:

1. Standardize terminology (e.g., DIC, not “DDIC”) and fix typographical errors that distract from the science.
2. Ensure figure/text alignment; Discussion cites “Figure 8” but only Figures 3–7 are present.
3. Harmonize tool names and panel labels across figures (e.g., “greyscale Stardist”), avoiding duplicates and inconsistent capitalization.
4. Abstract contains tool names, and language errors (e.g., “fluorescence in suit hybridization,” inconsistent “Startdist/NucleAizer”) which undermine clarity at first read.

20. Pearson correlation is used as proof of “supremacy” without justification; some coefficients/signs appear inconsistent or typographically erroneous (e.g., “-e0,839”).
21. Several figure descriptions lack scale bars, magnification, or ROI definitions alongside quantitative claims.

Overall, authors should be in a better position against existing evaluations and clearly state what is new beyond applying known tools to a small FL set. The methodological transparency, statistics, figure/text consistency, language, and validation strategy must be substantially improved.

Author Response

Thank you for your detailed views and suggestions to improve our manuscript!

Your can find our modifications to the original manuscript in red color.

Major issues:
The abstract is upgraded and supplemented with more specific data, as suggested.

We inserted a sentence on the diagnostic requirement of FISH analysis:
„FISH clinical guidelines for diagnostic application, depending of the type and genetic aberration tested
usually require the analysis of 50-200 tumor cells per sample (Wolff 2007; Mascarello 2011).”

For the rationale of method selection we inserted 2 paragraph into the revised manuscript:
Although several deep learning algorithms for spot detection are available (Bouilhol 2022; Xue 2023;
Jian 2024), their integration with nuclear segmentation frameworks remains essential, particularly
because only the signals localized within nuclei are diagnostically relevant (Wolff, 2007). The present
work focused pri-marily on nuclear detection and its clinical evaluation, as accurate identification of
nuclei is the prerequisite for reliable downstream FISH analysis. While deep neural networks provide
multiple options for spot de-tection, no unified framework is currently accessible that allows simultaneous
execution of nuclear and spot detection algorithms in an out of the box manner. To address this limitation,
a custom spot detection applica-tion has been under development, designed to identify signals specifically
within the masks of segmented nuclei across the spot channels (Kuczarov 2024).

The eye-control based method was choosen as a gold standard, since most diagnostic laboratories still
use traditional analysis of at least 50-200 tumor cells per sample depending on the type of tumor and gene
aber-ration required by the guidelines for a reliable diagnostic decision making (Wolff, 2007; Mascarello
2011). This fact and the frequent overlapping nuclei in lymphoma sections emphasized the need to search
for automation and compare the analytic performane in nuclear segmentation of FISHQuant and the AI-
based, general purpose CellPose, NucleAIzer and Stardist algorihms in order to increase the accuracy
and reproduciblitity of FISH reading reduce the chance for bias and fatigue of pathologists.
We agree that 5 cases tested with 2 different methods (10 samples) are low numbers, which we mention as
a limitation in the revised manuscipt. However, the detected 500-1000 nuclei per sample depending on the
method used on average 4000 nuclei per each method shold be enough for a valid statistical comparison.
Nevertheles, in the revised discussion we mention the case number as a potential limitation of our study.
We used off-he-shelf version without any further training method as stated in the Methods section of the
revised manuscript.

Manual annotation was performed by 2 pathologists based on images of consecutive H&E stained slides
proved to allow labeling representative areas of multilayer scanned FISH images (which approach is
close to 3D) for testing BCL2 gene fusion/translocation.
We also stated in theat the beginning of the Results, that:

The kappa correlation between 2 pathologists on the detected numer of cell nuclei using eye-contolled
examination with MarkerCounter was high (k=0,93186). Overall, the narrow ranges and moderate
deviations confirm that the manual quantification was reproduci-ble, with no evidence of extreme
discrepancies or systematic bias.
We also added a more precise definition both in the abstract and in the discussion about how average spot
numbers can validate nuclear segmentation:
The accuracy of nucleus detection by image processing algorithms is validated by the very similar
average number of gene spots per cell nuclei in the AI-based analysis to our assessors findings using eye
control (at high kappa values), despite the higher number of cell nuclei detected by AI algorithms. The
difference in nuclear count possibly explained by the left out gene spots we could localize unequivocally to
separate cell nuclei.
We inserted a sentence on the diagnostic requirements of FISH analysis:
FISH clinical guidelines for diagnostic application, depending of the type and genetic aberration tested
usually require the analysis of 50-200 tumor cells per sample (Wolff 2007; Mascarello 2011).

We also added a short paragraph to the revised manuscript on the upstear factors which can influence
FISH results:
„It must be emphasized that the effectiveness of any segmentation algorithm depends heavily on upstream
factors, such as high-quality sample preservation in formalin-fixed and paraffin embedded sections,
conditions of optimal FISH reaction, image resolution, signal-to-noise ratio and standardized imaging
protocols8. Poor sample quality or imaging artifacts can severely impact segmentation accuracy,
regardless of the algorithm used. It should also consider fluorophore choice, autofluorescence, clearing
chemistry, detector spectra bias segmentation and spot-calling; the study by Kumar, V. et al. 2021, Brain
Struct Funct 226, 481–499, shows far-red dyes (e.g., AF-647) consistently yield higher S/N than AF-
488/594, with weaker channels further degraded by clearing solvents and instrumentation. Routine
quality control steps, including normalization, denoising and artifact removal must therefore be integral
to any digital pathology workflow that seeks to take advantage of AI-powered tools9.”
The ethical approval issuing was corrected to 10/10/22

Minor issues:
The mistypins and abreviations were corrected in the revised manuscript.
Also, we improved the description of selection strategy on ROI, inserted scale bars to define object sizes
more clearly. Improved the language quality and figure text consistency (standardized abbreviations) in
the revised manuscript.
ROI (region of interest) areas were selected by hematopathologists (LK and BT) based in follicle-like
structures of frequent overlapping nuclei to include at least 1000 cells each, excluding areas of
autofluorescing red blood cell.
Mistyped Figure 8 was eliminated.

A new, Table 5. was inserted.

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

The revision substantially improves the manuscript; however, major issues remain and warrant another round, focused on endpoint clarity, statistical rigor, gold-standard definition, reproducibility, and real-world robustness, with earlier minor items (terminology, figure alignment, administrative details) consolidated into a single cleanup task.

Major remaining issues-

1. Please perform a full copy-edit to fix typos and standardize tool/technical terms (e.g., DIC, consistent StarDist/NucleAIzer naming). Also, ensure figure/text references exist and match (no missing “Figure 8”) and include scale bars/ROIs.
2. Ethics approval date still appears malformed in v2 (“issued in 10/10/205”), despite the response claiming it was corrected to 10/10/22; the manuscript must match the response. Synchronize the ethics approval date and any admin details across text/tables/response.
3. Authors should define one primary endpoint (e.g., per-case F1 for genetic class) with clearly labeled secondary metrics. Also, include confidence intervals (and, if feasible, simple bootstrap) for primary/secondary metrics; add per-case confusion matrices (aggregate in main, full in supplement). Please clean the correlation matrix (remove odd notations like “-e0,839”), drop the “negative r ⇒ superiority” language and benchmark VP/VS against standard overlap metrics (e.g., Dice/Jaccard).
4. Describe readers, blinding, and inter-rater agreement; add a blinded micro-panel adjudication for overlapping nuclei to test whether AI splits are biologically plausible.
5. For internal consistencies, reconcile statements where FISHQuant is “most reliable” vs AI “superior”, and fix any impossible values (e.g., F-measure >1) and ensure claims in Discussion match Results.
6. To further improve reproducibility and availability, provide a minimal availability package: exact configs/thresholds, seeds/model IDs, and either trained weights or a small reproducible notebook for the spot-decoder on example tiles.

Minor issues-

7. Author should state a clear code/data sharing plan (what will be public at acceptance).
8. Briefly acknowledge that fluorophore/autofluorescence/clearing/detector factors modulate S/N and can bias segmentation/spot-calling. If space allows then add a short sensitivity/robustness note for per-slide S/N reporting, cross-channel checks, and any depth or background controls.
9. For Limitations, add a concise list in main text addressing- small single-center sample, constrained gold standard, no external scanner/probe validation, parameterization incompleteness, potential selection bias, limited inter-reader analysis (beyond κ). Furthermore, point to how items in #3–#7 (above) mitigate or frame these limits.
10. In the Future directions paragraph, please outline next steps: multi-center and multi-scanner/probe validation, blinded panel ground truth per nucleus, comparisons with additional 3D models, preprocessing ablations, and full code/data release.
11. Finally, for Conclusions’ tone and scope, temper claims to match the evidence and revised statistics; avoid generalizing to “scalable clinical workflows” until external validation is shown.

Author Response

Dear Editor,

Thank you for your comments and suggestions which can further improve our manuscript!

Below, please find our point-by-point responses and revisions (in green)!

Major remaining issues:

1. We standardized our terms and applied scale bars to our morphology based figures. There is no Figure 8. It was a fragment from an earlier version, therefore we just erased the related sentence.
2. 10/10/22 is in the revised manuscript.
3. By analyzing VS and VP, we could systematically assess the reliability and generalizability of AI-based nuclei segmentation methods, ensuring their suitability for complex biological image analysis.  We put this sentence in to the introduction: “In this study, the primary endpoint was the per‑case F1‑score for correct genetic classification (fusion, break‑apart, normal), while all segmentation‑level metrics were treated as secondary endpoints.” Key evaluation metrics including Recall, Precision, F-measure and Accuracy also offer insights into algorithm performance…
   
   Odd notations and notes were corrected, the Dice and Jaccard metrix were also counted and added to the new Table 4.

Accordingly the new legend for Table 4.: “ Summary of the genetic classification abilities of the image analysis methods including Dice and Jaccard metrix in follicular lymphoma samples after FISH-reaction”

5. On reader, blinding, and inter‑rater agreement, we added sentences to the relevant sections of the Materials and methods section.
Digital FISH slides were independently reviewed by two board‑certified hematopathologists who were blinded to the algorithmic outputs. Inter‑rater agreement was quantified using Cohen’s kappa statistics. To adjudicate cases with overlapping or densely packed nuclei, a blinded micro‑panel of three additional pathologists was convened. This panel evaluated a representative subset of nuclei where AI‑based segmentation suggested nuclear splits, and reached consensus on whether the proposed separations were biologically plausible. The adjudication results were incorporated into the validation dataset to ensure that AI‑driven nuclear segmentation was consistent with expert biological interpretation.

6. We also corrected or erased the performance overestimating expressions such as “most reliable” and “superior”. We did not find F measure value which exceeded 1.

7. To further improve reproducibility and availability, provide a minimal availability package: exact configs/thresholds, seeds/model IDs, and either trained weights or a small reproducible notebook for the spot-decoder on example tiles.

Minor issues:

8. Here we can speak only on spot detection code, which however, is developed at 3DHistech and under industrialization. All others were 3^rd party applications.  

9. The quality of FISH reactions is strongly influenced by the laboratory preparation of paraffin blocks, the labeling technique, and the choice of fluorophores as it is already discussed. Soma added sentences also highlight that FITC may produce autofluorescence in paraffin‑embedded samples, which can bias signal detection. Fluorescent whole‑slide digitization systems incorporate automatic image enhancement autofluorescence filtering solutions, which in turn affect the visibility of gene signals. In addition, detector settings all modulate the signal‑to‑noise ratio and may impact segmentation and spot‑calling performance.

10. The limitation session is supplemented with this sentence as suggested:” Further limitations may involve the use of single-center samples, constrained gold standard, no external scanner/probe validation, parameterization incompleteness, potential selection bias, limited inter-reader analysis (beyond κ).”

So as the future directions as: “Future directions should focus on multi-center and multi-scanner/probe validation, blinded panel ground truth per nucleus, comparisons with additional 3D models, preprocessing ablations, and full code/data re-lease.”

Finally, In the revised manuscript we have moderated the tone of the Conclusions to ensure that our claims are fully aligned with the presented evidence and statistical results. Specifically, we removed references to “scalable clinical workflows” and instead emphasized that our findings demonstrate improved accuracy and reproducibility of nuclear segmentation in follicular lymphoma samples under controlled experimental conditions. We now clearly state that external validation in independent cohorts will be required before clinical scalability can be claimed. The revised Conclusions highlight the potential of AI‑based 3D segmentation methods while acknowledging the current limitations and the need for further validation.

An additional Table 5 was to be inserted (it was already mentioned in the text) but I can not see in the             R1 manuscript despite it was added as an extra table, now is provided with the relevant table text.