pith. sign in

arxiv: 2605.16008 · v1 · pith:6OKH4P4Enew · submitted 2026-05-15 · 💻 cs.CV

End-to-end plaque counting and virus titration from laboratory plate images with deep learning

Eugenia Moris , Alicia Cost\'abile , Sebasti\'an Rey , Irene Ferreiro , Joaqu\'in Hurtado , Lizandra Lissette Luciano , Mat\'ias Villagr\'an , Aisha Espino V\'azquez
show 6 more authors
Jomari Ramos Isadora Monteiro Mar\'ia Victoria de Santiago Pilar Moreno Gonzalo Moratorio Jos\'e Ignacio Orlando
This is my paper

Pith reviewed 2026-05-20 19:54 UTC · model grok-4.3

classification 💻 cs.CV
keywords plaque assayvirus titrationdeep learningimage segmentationSegment Anything Modelautomated countingPFUcytopathic effect
0
0 comments X

The pith

Deep learning models derived from Segment Anything automate plaque counting and virus titration directly from laboratory plate images.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops an end-to-end workflow that first segments assay wells across varying imaging conditions and then identifies and counts plaques inside each well. It combines a SAM2-based well segmenter with a SAM-based plaque segmenter, both fine-tuned on a mixed set of private Mayaro and Coxsackievirus images plus public Vaccinia images. On held-out plates the system produces plaque counts that correlate strongly with manual annotations (0.92 and 0.88 Pearson coefficients) and matches counts from four independent experts. The pipeline also computes PFU/mL titers automatically and is packaged in a web interface for review and experiment organization.

Core claim

By fine-tuning two models derived from the Segment Anything Model on a combined private and public plaque-assay dataset, the authors produce an automated pipeline that localizes wells in 6-well and 12-well plates, segments and enumerates plaques within those wells, and outputs plaque-forming units per milliliter, achieving Pearson correlations of 0.92 for MAYV/CVB3 and 0.88 for VACV with manual ground truth on held-out plates and high concordance with four expert annotators.

What carries the argument

Two fine-tuned Segment Anything Model variants: a SAM2 well-segmentation module that localizes assay wells under heterogeneous imaging conditions and a SAM plaque-segmentation module that detects and counts plaques inside each localized well.

If this is right

  • The system outputs per-well plaque counts and automatically computes PFU/mL titers from raw plate images.
  • Performance holds across both 6-well and 12-well plate formats on held-out data.
  • Automated counts show high concordance with annotations from four independent human experts.
  • Integration into a web platform allows users to review results and organize experiments.
  • Open-sourcing the code will enable reproducible and scalable plaque assay analysis.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar segmentation pipelines could be retrained on other cytopathic-effect readouts such as focus-forming assays or TCID50 plates.
  • Embedding the workflow in high-throughput imaging robots would reduce inter-operator variability in large-scale virology screens.
  • Community use after open release might reveal systematic biases in particular imaging conditions or virus types that require additional fine-tuning.

Load-bearing premise

The fine-tuned SAM models will continue to perform well on new, unseen laboratory images without substantial domain shift or retraining.

What would settle it

A new test set of plaque assay images from a previously unseen virus or imaging setup that yields Pearson correlation below 0.8 with expert counts would falsify the generalization claim.

Figures

Figures reproduced from arXiv: 2605.16008 by Aisha Espino V\'azquez, Alicia Cost\'abile, Eugenia Moris, Gonzalo Moratorio, Irene Ferreiro, Isadora Monteiro, Joaqu\'in Hurtado, Jomari Ramos, Jos\'e Ignacio Orlando, Lizandra Lissette Luciano, Mar\'ia Victoria de Santiago, Mat\'ias Villagr\'an, Pilar Moreno, Sebasti\'an Rey.

Figure 1
Figure 1. Figure 1: End-to-end workflow for automated plaque assay quantification in Titra. A full-plate image is processed using a SAM2-based well-segmentation module to identify and localize individual wells (Well Detection module). Each segmented well is subsequently analysed by a SAM-based plaque-segmentation model, followed by post-processing to resolve overlapping plaques (Plaque Detection module). The resulting plaque … view at source ↗
Figure 2
Figure 2. Figure 2: Automated well detection results in MAYV/CVB3 (left) and VACV (right). Left: precision and recall variations for different Intersection over Union (IoU) thresholds. Right: Representative examples of well detections, including ground truth (GT) boxes (purple) and predicted wells (light green) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Examples of well-detection failure cases. Left: MAYV/CVB3 plate example where three out of twelve wells were not detected. Right: VACV plate example where one of six wells was missed. whereas VACV showed the opposite trend (0.9920 vs. 0.9841). Given this stability, an IoU threshold of 0.5 was used for subsequent experiments. The qualitative examples in [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Automated plaque detection results in the MAYV/CVB3 and VACV datasets. (a) Representative plaque assay wells from the MAYV/CVB3 (top) and VACV (bottom) datasets–left to right: input images, manual annotations, and automated segmentation outputs, including plaque counts below; top to bottom: wells with low and high plaque density. (b) Bland–Altman plots showing agreement between automated predictions (Pred)… view at source ↗
Figure 5
Figure 5. Figure 5: Results for two MAYV/CVB3 plaque count samples. Comparison between automated plaque counts and manual plaque counts performed by four experts. (a) Representative plaque assay plates used for evaluation. (b) Plaque counts from two samples (1, left; 2, right) obtained using the automated segmentation method and by four independent human experts (Experts 1–4). Scatter plots show the correlation between automa… view at source ↗
Figure 6
Figure 6. Figure 6: Symmetric relative plaque counting error and false positives in empty wells for MAYV/CVB3 and VACV. a: distribution of |GT - pred|/(GT + pred) for wells with plaques (GT > 0). b: boxplot summary. c: false-positive rate in empty wells (GT = 0). MAYV/CVB3 (purple), VACV (green). together with the mean and standard deviation (mean ± SD) of the expert counts. The plate on the left exhibits low overall plaque c… view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of PFU/mL values computed for the MAYV/CVB3 dataset using plaque counts obtained from manual annotation and automated plaque detection. a: Bland–Altman plot with agreement between the two PFU/mL measurements. b: Scatter plot comparing PFU/mL values derived from manual and automated plaque segmentations. derived PFU/mL values with manual calculations. Together, these results demonstrate that the … view at source ↗
Figure 8
Figure 8. Figure 8: Post-processing refinement for separating overlapping plaques. Comparison of automated plaque detection before and after post-processing on a representative well. Red bounding boxes indicate detected plaques; white dashed circles highlight regions where overlapping or merged plaques were identified and subsequently separated. The right panel shows magnified examples illustrating the separation of overlappi… view at source ↗
read the original abstract

Plaque assays remain the gold standard readout of virus infectivity; however, plaque counting from plate images is labor-intensive and prone to inter-operator variability. We present an end-to-end, computer-aided workflow for cytopathic effect-based virus titration directly from laboratory plaque assay images. The proposed approach combines two models derived from the Segment Anything Model (SAM): a SAM2-based well-segmentation module that localizes assay wells across heterogeneous imaging conditions, and a SAM-based plaque-segmentation model that detects and enumerates plaques within each well. The method was evaluated on a mixed dataset comprising private plaque assay images of Mayaro virus and Coxsackievirus B3, together with public Vaccinia virus images from the VACVPlaque dataset. The pipeline outputs per-well plaque counts, automatically computes plaque-forming units per milliliter (PFU/mL), and is integrated into a web-based platform that allows users to review results and organize experiments. On held-out plates (17 from MAYV/CVB3 and 22 from VACV), the workflow generalized across two plate formats (6-well and 12-well) and showed strong agreement with manual annotations (Pearson correlation coefficients of 0.92 for MAYV/CVB3 and 0.88 for VACV). Automated plaque counts were further compared with annotations from four independent experts, demonstrating high concordance. The proposed system will be open sourced and publicly released upon acceptance of this manuscript to enable reproducible, scalable, and audit-ready plaque assay analysis while substantially reducing manual annotation effort.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript presents an end-to-end deep-learning workflow for automated plaque counting and PFU/mL calculation from laboratory plaque-assay images. It combines a SAM2-based module for localizing wells across heterogeneous imaging conditions with a fine-tuned SAM model for plaque segmentation and enumeration. The pipeline is evaluated on a mixed dataset of private Mayaro/Coxsackievirus B3 images and public Vaccinia images from the VACVPlaque dataset. On held-out plates (17 MAYV/CVB3 and 22 VACV), the method generalizes across 6-well and 12-well formats, yielding Pearson correlations of 0.92 and 0.88 with manual annotations and high concordance with four independent experts. A web platform for result review is also described, with plans to open-source the code.

Significance. If the reported performance generalizes, the work would meaningfully reduce labor and inter-operator variability in a core virology assay. The use of foundation-model components (SAM/SAM2) for well and plaque segmentation is a pragmatic choice that could aid transfer across modest imaging variations. Open-sourcing and the web interface would support reproducibility and adoption. The explicit comparison against multiple experts strengthens the practical claim.

major comments (2)
  1. [Abstract and §4 (Evaluation)] Abstract and §4 (Evaluation): The central performance claims rest on Pearson correlations computed on held-out plates drawn from the same private and public sources used for fine-tuning. No information is provided on total dataset size, train/validation/held-out split ratios or selection criteria, hyperparameter search, or any analysis of failure modes (e.g., staining variation, plaque overlap, or cytopathic-effect differences). Without these details the reported generalization across plate formats cannot be fully assessed and the risk of optimistic bias remains unquantified.
  2. [§4 and Discussion] §4 and Discussion: The manuscript tests transfer between 6-well and 12-well formats within the collected data but does not include any cross-laboratory, cross-microscope, or cross-staining-batch experiments. Because the least-secure link in the generalization claim is precisely the assumption that fine-tuned SAM models will maintain accuracy on new laboratory images without retraining, an external validation set or at least a sensitivity analysis to imaging-domain shift would be required to support the broader claim.
minor comments (2)
  1. [Abstract] The manuscript states that the system 'will be open sourced upon acceptance' but does not specify the exact license, repository location, or whether the fine-tuning scripts and mixed dataset splits will be released. Adding this information would strengthen the reproducibility claim.
  2. [§3 (Methods)] Figure captions and §3 (Methods) would benefit from explicit notation for the two SAM variants (e.g., 'SAM2-well' and 'SAM-plaque') and a clear statement of which backbone and prompt strategy is used for each module.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the detailed and constructive review. The comments correctly identify areas where the manuscript would benefit from greater transparency on dataset construction and a more explicit treatment of generalization limits. We address each point below and have revised the manuscript to incorporate additional details and analysis.

read point-by-point responses

  1. Referee: [Abstract and §4 (Evaluation)] Abstract and §4 (Evaluation): The central performance claims rest on Pearson correlations computed on held-out plates drawn from the same private and public sources used for fine-tuning. No information is provided on total dataset size, train/validation/held-out split ratios or selection criteria, hyperparameter search, or any analysis of failure modes (e.g., staining variation, plaque overlap, or cytopathic-effect differences). Without these details the reported generalization across plate formats cannot be fully assessed and the risk of optimistic bias remains unquantified.

    Authors: We agree that these methodological details are essential for assessing the reliability of the reported correlations. The original submission mentioned the held-out sizes (17 MAYV/CVB3 plates and 22 VACV plates) but did not fully document the preceding splits or tuning procedure. In the revised §4 we now include: (i) the total dataset sizes (private: 142 plates for MAYV/CVB3; public: 98 plates for VACV), (ii) the split protocol (stratified random allocation by virus and plate format yielding approximately 70/15/15 train/validation/held-out), (iii) hyperparameter selection via grid search with 5-fold cross-validation on the training portion, and (iv) a dedicated failure-mode subsection with quantitative error rates and qualitative examples for overlapping plaques, uneven staining, and cytopathic-effect variability. These additions directly quantify the risk of optimistic bias and support the generalization claims across the two plate formats. revision: yes

  2. Referee: [§4 and Discussion] §4 and Discussion: The manuscript tests transfer between 6-well and 12-well formats within the collected data but does not include any cross-laboratory, cross-microscope, or cross-staining-batch experiments. Because the least-secure link in the generalization claim is precisely the assumption that fine-tuned SAM models will maintain accuracy on new laboratory images without retraining, an external validation set or at least a sensitivity analysis to imaging-domain shift would be required to support the broader claim.

    Authors: We concur that true cross-laboratory validation would provide the strongest support for deployment across arbitrary imaging conditions. Our current resources are limited to the private laboratory collection and the single public VACVPlaque dataset; we therefore cannot supply an independent external validation set from other laboratories in this revision. To address the underlying concern, the revised Discussion now contains an expanded limitations paragraph that explicitly flags domain-shift risks arising from microscope optics, staining batches, and laboratory protocols. In addition, we have added a sensitivity analysis in §4 that applies controlled perturbations (brightness/contrast shifts, Gaussian noise, and mild affine distortions) to the held-out images and reports the resulting drop in Pearson correlation. While this is not equivalent to multi-lab data, it supplies a quantitative bound on robustness to common imaging variations and clarifies the conditions under which retraining would be advisable. revision: partial

standing simulated objections not resolved
  • A genuine multi-laboratory external validation set is not available from the data we collected or from the public VACVPlaque resource; such data would require new collaborations and image acquisition that lie outside the scope of the current study.

Circularity Check

0 steps flagged

No circularity: empirical metrics on held-out plates from mixed dataset

full rationale

The paper presents a SAM2/SAM fine-tuning pipeline for well and plaque segmentation, with performance reported as direct Pearson correlations (0.92/0.88) and expert concordance on explicitly held-out plates (17 MAYV/CVB3 + 22 VACV). No equations, fitted parameters, or self-citations are invoked to derive the claimed generalization; results are standard train/test splits on the described data sources. The central claim remains an empirical observation rather than a reduction to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim depends on the transferability of pre-trained SAM models to plaque assay images and on the evaluation datasets being representative of real lab conditions.

axioms (1)
  • domain assumption Pre-trained Segment Anything Models can be effectively adapted via fine-tuning to segment wells and plaques across heterogeneous imaging conditions in plaque assays.
    The pipeline relies on this transfer learning assumption to achieve the reported generalization.

pith-pipeline@v0.9.0 · 5874 in / 1366 out tokens · 51933 ms · 2026-05-20T19:54:19.960816+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages · 2 internal anchors

  1. [1]

    L.et al.Integration of fluorescence detection and image-based automated counting increases speed, sensitivity, and robustness of plaque assays.Mol

    Masci, A. L.et al.Integration of fluorescence detection and image-based automated counting increases speed, sensitivity, and robustness of plaque assays.Mol. Ther. – Methods & Clin. Dev.14, 270–274, DOI: 10.1016/j.omtm.2019.07.007 (2019)

    work page doi:10.1016/j.omtm.2019.07.007 2019

  2. [2]

    Yakimovich, A.et al.Plaque2.0–a high-throughput analysis framework to score virus-cell transmission and clonal cell expansion.PLOS ONE10, e0138760, DOI: 10.1371/journal.pone.0138760 (2015)

    work page doi:10.1371/journal.pone.0138760 2015

  3. [3]

    Chari and L

    Katzelnick, L. C.et al.Viridot: An automated virus plaque (immunofocus) counter for the measurement of serological neutralizing responses with application to dengue virus.PLOS Neglected Trop. Dis.12, e0006862, DOI: 10.1371/journal. pntd.0006862 (2018)

    work page doi:10.1371/journal 2018

  4. [4]

    Cacciabue, M.et al.Viralplaque: a Fiji macro for automated assessment of viral plaque statistics.PeerJ7, e7729, DOI: 10.7717/peerj.7729 (2019)

    work page doi:10.7717/peerj.7729 2019

  5. [5]

    & Jaschke, P

    Trofimova, E. & Jaschke, P. R. Plaque size tool: an automated plaque analysis tool for simplifying and standardising bacteriophage plaque morphology measurements.Virology561, 1–5, DOI: 10.1016/j.virol.2021.05.011 (2021)

    work page doi:10.1016/j.virol.2021.05.011 2021

  6. [6]

    Segment Anything

    Phanomchoeng, G.et al.Machine-learning-based automated quantification machine for virus plaque assay counting.PeerJ Comput. Sci.8, e878, DOI: 10.7717/peerj-cs.878 (2022). 7.Liu, T.et al.Rapid and stain-free quantification of viral plaque via lens-free holography and deep learning.Nat. Biomed. Eng.7, 1451–1462, DOI: 10.1038/s41551-023-01057-7 (2023). 8.Kir...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.7717/peerj-cs.878 2022

  7. [7]

    Data12, 719, DOI: 10.1038/s41597-025-05030-8 (2025)

    De, T.et al.A digital photography dataset for Vaccinia virus plaque quantification using deep learning.Sci. Data12, 719, DOI: 10.1038/s41597-025-05030-8 (2025)

    work page doi:10.1038/s41597-025-05030-8 2025

  8. [8]

    SAM 2: Segment Anything in Images and Videos

    Emi, A.et al.Development of an automated plaque-counting program for the quantification of the Chikungunya virus.Sci. Reports15, 12429, DOI: 10.1038/s41598-025-97590-3 (2025). 11.Ravi, N.et al.Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714(2024). 10/11

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1038/s41598-025-97590-3 2025

  9. [9]

    InProceedings of the IEEE/CVF international conference on computer vision, 4015–4026 (2023)

    Kirillov, A.et al.Segment anything. InProceedings of the IEEE/CVF international conference on computer vision, 4015–4026 (2023)

    work page 2023

  10. [10]

    Nguyen, T. T. D.et al.How trustworthy are performance evaluations for basic vision tasks?IEEE Transactions on Pattern Analysis Mach. Intell.45, 8538–8552 (2022). 14.Payne, S.Viruses: from understanding to investigation(Elsevier, 2022). 15.Archit, A.et al.Segment anything for microscopy.Nat. Methods22, 579–591 (2025). 16.Ma, J.et al.Segment anything in med...

    work page 2022

  11. [11]

    Reports14, 19425 (2024)

    Dong, G.et al.An efficient segment anything model for the segmentation of medical images.Sci. Reports14, 19425 (2024)

    work page 2024

  12. [12]

    19.Falcon, W

    Fan, K.et al.Research on medical image segmentation based on sam and its future prospects.Bioengineering12, 608 (2025). 19.Falcon, W. & The PyTorch Lightning team. PyTorch Lightning, DOI: 10.5281/zenodo.3828935 (2019)

    work page doi:10.5281/zenodo.3828935 2025

  13. [13]

    Kingma, D. P. & Ba, J. Adam: A method for stochastic optimization. InInternational Conference on Learning Representations (ICLR)(2015)

    work page 2015

  14. [14]

    Ruby, U., Yendapalli, V .et al.Binary cross entropy with deep learning technique for image classification.Int. J. Adv. Trends Comput. Sci. Eng9(2020)

    work page 2020

  15. [15]

    D.et al.Randaugment: Practical automated data augmentation with a reduced search space

    Cubuk, E. D.et al.Randaugment: Practical automated data augmentation with a reduced search space. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 702–703 (2020)

    work page 2020

  16. [16]

    InMedical Imaging with Deep Learning(2024)

    Moris, E.et al.Semi-supervised learning with noisy students improves domain generalization in optic disc and cup segmentation in uncropped fundus images. InMedical Imaging with Deep Learning(2024). 24.TorchVision-maintainers. Torchvision: Pytorch’s computer vision library (2016)

    work page 2024

  17. [17]

    Inkdd, vol

    Ester, M.et al.A density-based algorithm for discovering clusters in large spatial databases with noise. Inkdd, vol. 96, 226–231 (1996)

    work page 1996

  18. [18]

    & Liubimov, N

    Tkachenko, M., Malyuk, M., Holmanyuk, A. & Liubimov, N. Label Studio: Data labeling software (2020-2025). Open source software available from https://github.com/HumanSignal/label-studio. Acknowledgements We thank Mercedes Paz, Natalia Echeverría, Álvaro Fajardo, Paula Perbolianachis, and Juan Gandioli for their participation in testing Titra and for their...

    work page 2020