Data Augmentation-Driven Improvements in Malignant Lymphoma Image Classification
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsSummary:
This work applies data augmentation to a publicly available digital pathology dataset in the aim of more correctly identifying pathology than prior work. The authors give a brief but relevant history of some relevant research in this area. They explain the difficulties with collecting large amounts of data for medical computer vision problems. Their approach, used by many, is to do data augmentation to artificially increase the amount of data available for training a machine learning model. Specifically, the authors 'slice' the data much like patch selection or tiling. The Convolutional Neural Network approach used by the researchers performed better with their approach to data augmentation + slicing compared to either no augmentation + slicing or no augmentation. There are some minor errors or improvements to make. The figures and tables are relevant and the references are well chosen. They conclude by saying that there are more advanced methods available today with an implication that this CNN approach may no longer be relevant.
Opinion:
In terms of novelty this isn't a very flashy report. That's ok. I feel that the slice method is very similar to the concept of patches or tiling that are already widely in use for training deep learning models, including CNN's. The report would be more powerful if the authors shared their code. It is nice that the data are publicly available. I wasn't clear about their sentiment in concluding paragraphs about more sophisticated models being available at the time of submission.
Strengths:
Well supported work with clear tables and figures. Acronyms and metrics are defined. Uses a public dataset. Cites what is relevant.
Weaknesses:
Code is not shared. Possible error in Table 6. New method is similar to existing and commonly used methods in computer vision. Unclear how this method generalizes to other problems. Unclear how much of a computational impact it is to get this improvement. Some ambiguous comparisons in conclusion.
Areas for Improvement:
I am accustomed to describing the work in a paper in the past tense and knowledge represented in the literature in present tense (or past tense if that is more relevant).
Some complex sentences use only commas to separate ideas. It can be helpful to reduce how many commas are used or use dashes or semicolons for better readability. For example:
"In addition to just splitting images, dataset augmentation, using geometric augmentations, is also performed."
can be
"Beyond image splitting, dataset and geometric augmentation were performed."
Though the data acquisition was described in the source publication, it would be good to repeat the instrument used to collect data including the camera. From the source publication: "Ten different cases of three different lymphomas (CLL, FL,
MCL; 30 slides total) were imaged on a Zeiss Axioscope white
light microscope with a 20x objective and a color CCD camera
AxioCam MR5. The slides were imaged with the same instru-
ment settings and same objective lens, camera, and light source.
Therefore, no other normalization was performed for the camera
channels."
Line Item changes:
108 - "RQ" must mean 'Researcher Question' but it is not defined. I would just list as 'question 1,' 'question 2', ...
348 - same as prior comment
128 - replace period in "(CLL)." with a comma "(CLL),"
123 - please mention the database source since the publication does not mention this. I know it is in the end of the manuscript, but I had to search for it.
https://www.kaggle.com/datasets/andrewmvd/malignant-lymphoma-classification/
Figure 5 - not labeled
Figure 5 - the images or captions for 'flip horizontal' and 'flip vertical' appear to have been switched
Table 5 - I find the min/max/avg columns confusing. I am not used to seeing different metrics summarized like this.
Table 5 - Also, I wonder if the difference scores were accidentally duplicated because they are the same for all three pairwise comparisons. It's hard to tell if this is intentional, but it seems to differ from Table 4, Fig 6, and the conclusions.
366-368 - what kind of state-of-the-art techniques are you referring to? VTT's? Please be clear and cite references or at least a review. In what way do you think they are more advanced?
Author Response
Respected reviewer,
thank you for your assistance in reviewing our manuscript. Please find the responses to your questions attached.
Kind regards,
the authors.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis manuscript presents a well-structured and empirically sound study on the effect of data augmentation in improving CNN-based classification of malignant lymphoma histopathological images. The authors use a relatively small public dataset (374 high-resolution images) and apply a two-tier preprocessing strategy: (1) systematic image slicing into 224×224 patches, and (2) deterministic geometric augmentation (rotations and flips). The study convincingly demonstrates that image slicing alone yields a substantial performance boost (e.g., accuracy from 0.73 to 0.94), while the combination of slicing and augmentation pushes classification performance above 0.95 for all reported metrics (accuracy, precision, sensitivity, specificity, AUC, and F1 score). The results are thoroughly benchmarked against existing literature, and the authors clearly articulate that the gains are achieved using a relatively simple architecture (VGG16) without reliance on synthetic data generation or advanced deep models like transformers.
Compared to similar studies, e.g., El Achi et al. (2019), Tambe et al. (2019), or Soltane et al. (2022), this paper’s approach stands out for its practical emphasis on computationally inexpensive methods that improve performance to or above state-of-the-art levels without requiring high-end architectures. One minor limitation is that the discussion, author does not explore, whether the same augmentation strategy generalizes across other histological image datasets or modalities (e.g., CT, MRI). Additionally, while slicing and augmentation improve performance, they significantly inflate training data size and computational cost, a tradeoff worth deeper discussion for real-world deployment. Given the comprehensive experimental validation, clarity of exposition, and practical value of the findings, I recommend this manuscript for minor revision. Suggested revisions include a short additional analysis or discussion of (1) computational resource tradeoffs versus augmentation gains and (2) potential applicability of the pipeline to other disease domains. Including experiments with more modern architectures or hybrid models (even briefly) could strengthen the broader applicability claims.
Author Response
Respected reviewer,
Thank you for providing comments on our manuscript. Please find our responses below. Any text added thanks to these comments was marked with purple text.
- computational resource tradeoffs versus augmentation gains and
The following text was added:
“It may be pertinent to discuss trade-offs between score improvements and the computational cost of adding the slicing and augmentation step. While the slicing and augmentation are relatively simple computational tasks — especially if discussing applying it on a single image for inference, they are essentially negligible. The largest impact on the performance comes from the large increase in the amount of images used for training. Still, this increase in computational time is kept to the training process which is completed once, prior to implementation, and it would not have significant impact during the exploitation of the developed models.”
- potential applicability of the pipeline to other disease domains. Including experiments with more modern architectures or hybrid models (even briefly) could strengthen the broader applicability claims.
The following text was added to the end of the Conclusion:
“One of the key issues that need to be addressed prior to focusing on the applicability of the suggested methodology is identifying other diseases where images collected are significantly larger than the CNNs are designed for, and that may benefit from the image slicing and augmentation techniques shown here. This would allow for more detailed conclusions relating to the applicability of the augmentation methodology provided in this paper to different diseases”
Kind regards,
the authors
Reviewer 3 Report
Comments and Suggestions for AuthorsThis study investigates the impact of a new augmentation approach on the classification performance of malignant lymphoma histopathological images, which involves slicing high-resolution images (1388×1040 pixels) into smaller segments (224×224 pixels) prior to applying standard augmentation techniques such as flipping and rotation. Experimental results also demonstrate a substantial improvement in classification performance when slicing is employed, with additional, albeit smaller, gains achieved through subsequent augmentation.
(1) The motivation of this paper should enhanced. Sinch there are too many works focusing on the data augmentation, the authors should describe the shortcomings of these methods, and how does these shortcomings promote the creation of the new method?
(2) The framework of your method should be included in the section 2 (Methods and Materials).
(3)Future work should be included to enhance your conclusion.
[1] Xiang S, Qian D, Guan M, et al. Less is more: Learning from synthetic data with fine-grained attributes for person re-identification[J]. ACM Transactions on Multimedia Computing, Communications and Applications, 2023, 19(5s): 1-20.
[2] Ran W, Yu Z, Xiang S, et al. A Training-Free Correlation-Weighted Model for Zero-/Few-Shot Industrial Anomaly Detection with Retrieval Augmentation[C]//ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025: 1-5.
Author Response
Respected reviewer,
thank you for your assistance in reviewing our manuscript. Please find the responses to your questions attached.
Kind regards,
the authors.
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsThe article presents an application of computer analysis of bioimages for the case of lymphoma. Though the technique is not ideal, it is good as prove of improvement for data augmentation.
Minor remarks – not need use extra abbreviations in the Abstract (line 8 and 9) – MCL, CCL and FL. The full names of the diseases are given, and abbreviations were used only once in the Abstract. But VGG16 abbreviation should be commented.
Line 16 – keywords. Add keyword ‘histopathology’.
Fix the typo ‘lymphoma)’
Line 49: ‘CT images’ – give CT abbreviation in full
typo – ‘ReducedFireNet.’
Table 1. State-of-the-art results – please give more precise title – “results for lymphoma prediction or like that”
Line 130: ‘VGG16 network’ – should be commented in detail, with reference.
This paragraph marked in color and not complete. What is ‘Figure ??’
Figure 3 – may make more simple pie-char or remove it. All the numbers are in the text.
Formulas (1-4) – with standards parameters are too common. Parameters like false negatives, true positives normally are given with Capital letters – TP, Tn and like that
Table should fit to the page width
Line 410 : ‘is also a planned future work’ – too standard concluding phrase. Just remove it, or rephrase.
Author Response
Respected reviewer,
Thank you for your review of our manuscript, please find the point-by-point comments addressed below. The changes made to the manuscript relating to these were marked with green text.
1. Minor remarks – not need use extra abbreviations in the Abstract (line 8 and 9) – MCL, CCL and FL. The full names of the diseases are given, and abbreviations were used only once in the Abstract. But VGG16 abbreviation should be commented.
The abbreviations in the abstract were adjusted according to the comment.
- Line 16 – keywords. Add keyword ‘histopathology’.
The keyword was added.
- Fix the typo ‘lymphoma)’
The erroneous bracket was removed.
- Line 49: ‘CT images’ – give CT abbreviation in full
The abbreviation was given in full first.
- typo – ‘ReducedFireNet.’
The text was corrected to “Reduced FireNet”
- Table 1. State-of-the-art results – please give more precise title – “results for lymphoma prediction or like that”
The title was changed to “State-of-the-art results in image-based lymphoma classification”
- Line 130: ‘VGG16 network’ – should be commented in detail, with reference.
The following text was added with citation: “VGG16 CNN (a deep convolutional neural network known for its simple architecture of stacked 3×3 convolutional layers and 16 weight layers, widely used for image classification tasks)” - This paragraph marked in color and not complete. What is ‘Figure ??’
We apologize, as the figure did not load while compiling for some reason. We have recompiled and the image is now correctly added and referenced.
- Figure 3 – may make more simple pie-char or remove it. All the numbers are in the text.
The image and the references to it were removed.
- Formulas (1-4) – with standards parameters are too common. Parameters like false negatives, true positives normally are given with Capital letters – TP, Tn and like that
Formulas were adjusted to use capital letters.
11.Table should fit to the page width
The table was adjusted so it’s visible in its entirety.
- Line 410 : ‘is also a planned future work’ – too standard concluding phrase. Just remove it, or rephrase.
Kind regards,
the authors
Round 2
Reviewer 3 Report
Comments and Suggestions for AuthorsThe authors have addressed all of my concerns, now this version can be acceptable in this reputated journal.
Author Response
We would like to thank the reviewer for their re-review of our manuscript.
Kind regards,
the authors
