List of Papers Browse by Subject Areas Author List
Abstract
Detecting BK Virus (BKV) is crucial for managing post-transplant outcomes in kidney patients. While BKV is typically identified using SV40 immunohistochemistry (IHC), this method is time-consuming, limited by tissue availability and resource-intensive, especially in low-resource settings. Recent advances in computational pathology have shown potential for automating disease detection from Hematoxylin and Eosin (H&E)-stained images, though BKV detection remains understudied due to its low prevalence and limited data. We hypothesize that BKV-positive cells exhibit unique morphological patterns in H&E-stained tissue that are detectable via computational methods. To address this, we developed BKVision, a weakly-supervised deep learning model for BKV detection in H&E whole-slide images (WSIs). Trained on 3,734 WSIs, BKVision achieves an F1-score of 0.984 ± 0.008 on a test cohort of 936 slides. Additionally, we conducted a morphological analysis on 774 H&E image patches, extracting 37 human-interpretable features and validating them against IHC with pathologist guidance. This identified 11 cell attributes, including nuclear enlargement and chromatin texture changes, that distinguish BKV-positive from negative cases. These findings highlight the potential to enhance BKV diagnostic criteria by integrating these identified morphological features. BKVision demonstrates the potential of computational methods to provide accurate, accessible, and interpretable BKV detection without requiring IHC, offering a cost-effective alternative in low-resource settings while revealing key morphological features of BKV infection.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/5442_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{SahSha_Automated_MICCAI2025,
author = { Sahai, Sharifa and Ramos-Guerra, Ana D. and Almagro-Pérez, Cristina and Jaume, Guillaume and Zhang, Andrew and Rennke, Helmut and Weins, Astrid and Ortuño, Juan E. and Ledesma-Carbayo, Maria J. and Mahmood, Faisal},
title = { { Automated Detection of BK Virus in H&E Whole-Slide Images Using Weakly-Supervised Deep Learning and Interpretable Morphological Biomarkers } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15967},
month = {September},
}
Reviews
Review #1
- Please describe the contribution of the paper
Researchers developed, a weakly-supervised deep learning model for detecting BK Virus (BKV) in Hematoxylin and Eosin (H&E)-stained images. The presented model (BKVision) achieved an F1-score of 0.984 on a test cohort, demonstrating its potential for accurate BKV detection. By analyzing 34 human interpretable features extracted from H&E image patches, the study identified 11 key morphological attributes of BKV-positive cells that distinguish them from negative cases. This proposed approach offers a cost-effective and accessible alternative to traditional SV40 immunohistochemistry, enabling low-resource settings to detect BKV without the need for resource-intensive IHC.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The method offers an alternative way to detect BKV from HE images rather than the time-consuming IHC (SV40)
The author compared the use of several different pretrained embedding models to achieve MIL-based classification task in BKV detection. They presented a quantitative comparison between methods.
Their results show strong performance, showing the possibility of BKV detection from HE images.
The authors not only demonstrated strong performance through their algorithm but also succeeded in showing an association between human interpretable morphological/cytological features and the algorithm findings (highlighted areas), grounding the model in human interpretable features.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The methodology used does not involve any technical novel. It is a pure application of well-established methods from the literature to a new problem.
Section 4.2 is the major portion of the paper’s contribution. However, the authors did not provide a quantitative analysis in this section. Figure 2 presents some analysis, but the details are not clear. A more detailed analysis is required.
It is not clear why the authors used a batch size of 1. In section 3.5, the authors also state that “Specifically, during training, we upsampled the BKV-positive class at the slide-level (n=130 slides) by adding more positive-case patches into each batch during training.” How is this possible if the batch size is 1? Do they mean they used a single patch in each iteration, or do they mean batch size of one slide per iteration?
It is not clear how these nuclear/morphological characterizations can be used by the users. Is the extraction of all these features automated? The authors mentioned that they used “HoVerNet” model and “Histocartography library” to extract the features. Does this mean, given a patch, all the 11 features can be extracted from them automatically? If that is so, then the question is, can these features be used for any further analysis supporting the finding of the CLAM model? Would the extraction of these features help in more borderline cases where BKV detection in HE is harder?
The authors state that “Nuclei from the top 30 high-attention patches associated with positive BKV cases”. This information is used to calculate the statistics for BKV+ cases. The authors did not explicitly state how patch selection is conducted to calculate BKV negative cases.
The author state that “Data and model weights can be provided upon institutional approval”. Therefore, the dataset and the algorithm are easily accessible to the community for reproducing the results and benefiting from the method.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Even if the presented method is a pure application of well known methods in the literature, the authors present a concrete set of experiments to show applicability and diagnostic performance on an important problem. Moreover, their effort in correlating algorithmin findings with human interpretable features is an important step in making the method more transparent. However important details regarding this step are missing
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper offers novel insights by tackling the underexplored task of BKV detection via H&E images, which is traditionally diagnosed with IHC. The approach of combining weakly-supervised deep learning with morphological interpretability adds both a technical and clinical contribution. The technical approach sounds solid: the dataset size is substantial, and performance metrics are reported with statistical variance.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
This paper could have significant clinical impact, especially in low- resource settings where access to IHC is limited. Additionally, the morphological findings have the potential to refine diagnostic criteria. The impact is both practical (cost-effective diagnosis) and scientific (identifying morphological markers).
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- Clarify the weak supervision method.
- Include more technical specifics on morphological validation.
- Consider briefly stating the clinical implications of integrating BKVision into workflow.
- Mention any limitations or areas for future work to show critical reflection.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper is highly relevant and focuses on computational pathology. It introduces a novel application of weakly-supervised learning for a clinically relevant and underserved problem — BKV detection in kidney transplant patients — which aligns with CAI themes. This paper introduces a novel application, high performance, large dataset, clinically relevant, interpretable model.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The authors applies standard methods for computational pathology (Cpath) to detecting BK Virus from H&E image alone. Postprocessing uses the attention map from H&E on IHC slide to select interesting regions that the clinicians can study. The classifier is tested on data collected internally, using train/val/test split.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- The motivation of the paper is clear and important.
- The approach is simple and the results convincing.
- The paper is overall well written and easy to understand.
- On top of quantitative results, the authors also show clinically relevant information and analysis of the results.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- Even though the application is novel and relevant, there is no real methodological novelty.
- The experiments lack the use of more recent foundation models (FM)
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper seems quite straightforward: a well-established method is applied to a new, important application; extensive ablation studies are shown; overall the paper is well structured. I reserve the use of “strong” accept because a methodological novelty is missing.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
Reviewer 1 – Clarify the weak supervision method. We employ weak supervision using slide-level labels without requiring patch-level annotations. The model assumes that BKV-positive slides contain some informative patches and learns to identify them by optimizing slide-level classification. This enables the model to infer local patch features from global diagnoses. – Include more technical specifics on morphological validation. We validated selected features through literature review and expert pathologist input, confirming alignment with known BKV traits (e.g., nuclear enlargement, hyperchromasia). Automatically quantified differences between infected and non-infected nuclei were consistent with clinical expectations, supporting the objectivity of our pipeline. – Consider briefly stating the clinical implications of integrating BKVision into workflow. BKVision can assist pathologists by predicting infection directly from HE slides,potentially eliminating the need for SV40 IHC, and highlighting informative patches and nuclear features (e.g., area, shape) that support diagnostic decisions. – Mention any limitations or areas for future work to show critical reflection. Future work includes validating the discovered morphological biomarkers in larger cohorts and extending BKVision to detect coexisting conditions like rejection, which are often misdiagnosed and lead to inappropriate treatment. Reviewer 2 – Lack of quantitative analysis. We agree that Section 4.2 could be more detailed. To quantify differences, we computed log₂ fold-changes between features in high-attention patches from BK(+) and BK(–) slides. This approach, inspired by gene expression analysis, highlights significant distributional shifts and enhances interpretability. – Clarification on batch size and upsampling. We use an attention-based MIL model where each slide (bag) contains a variable number of patches (instances). Due to this variability, batch size was set to 1. We upsampled BK(+) cases by increasing the frequency of positive slide sampling, not patches. – Clarification on feature extraction and usage. Our pipeline is fully automated. After slide-level classification, top 30 high-attention patches per slide are selected. HoVer-Net segments nuclei, and Histocartography extracts pathomic features, 11 of which were found significantly associated with infection. This pipeline provides objective morphological cues from H&E, even in the absence of reliable visual signals. Reviewer 3 – Lack of methodological novelty. Our contribution lies in integrating MIL-based slide classification with downstream automated pathomics to support clinical interpretation. This fusion enables both accurate prediction and human-interpretable morphological insight. – Lack of recent foundation model comparisons. We compare against three recent histopathology-specific foundation models: CTransPath (’21), CONCH (’24), and UNI-2h (’25). The inclusion of CONCH and UNI-2h, both released within the past year, ensures that our evaluation reflects current state-of-the-art. Future work will explore additional models as they emerge.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A