Abstract

Sparse autoencoders (SAEs) emerged as a promising tool for mechanistic interpretability of transformer-based foundation models. Recently, SAEs were also adopted for the visual domain, enabling the discovery of visual concepts and their patch-wise attribution to input images. While foundation models are increasingly applied to medical imaging, tools for interpreting their predictions remain limited. In this work, we propose CytoSAE, a sparse autoencoder trained on over 40,000 peripheral blood single-cell images. CytoSAE generalizes well to diverse and out-of-domain datasets-including bone marrow cytology. Here, it identifies morphologically relevant concepts which we validated with medical experts. Furthermore, we demonstrate scenarios in which CytoSAE can generate patient-specific and disease-specific concepts, enabling the detection of pathognomonic cells and localized cellular abnormalities at patch-level. We quantified the effect of concepts on a patient-level AML subtype classification task and show that CytoSAE concepts reach performance comparable to the state-of-the-art, while offering explainability on the sub-cellular level. Source code and model weights are available at https://github.com/dynamical-inference/cytosae.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4885_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/dynamical-inference/cytosae

Link to the Dataset(s)

MLL23: https://github.com/marrlab/MLL23 Acevedo: https://data.mendeley.com/datasets/snkd93bnjr/1 Matek19: https://doi.org/10.7937/tcia.2019.36f5o9ld BMC: https://doi.org/10.7937/TCIA.AXH3-T579 AML_Hehr: https://doi.org/10.7937/6ppe-4020

BibTex

@InProceedings{DasMuh_CytoSAE_MICCAI2025,
        author = { Dasdelen, Muhammed Furkan and Lim, Hyesu and Buck, Michele and Götze, Katharina S. and Marr, Carsten and Schneider, Steffen},
        title = { { CytoSAE: Interpretable Cell Embeddings for Hematology } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15973},
        month = {September},

}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper presents CytoSAE, a deep learning framework designed for the unsupervised extraction of interpretable concepts from the DinoBloom-B pretrained model across diverse H&E-stained histopathology datasets. The authors demonstrate that CytoSAE can effectively learn and represent concepts at multiple levels of granularity, including the cell, patient, and disease levels.

  • 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 paper provides a clear and well-structured introduction that effectively outlines the problem and situates the work within the context of existing literature.
    • Visualizations are thoughtfully designed, with particularly helpful example images illustrating the learned interpretable concepts, enhancing both clarity and interpretability of the results.
    • The use of multiple datasets underscores the generalizability and broad applicability of CytoSAE across diverse settings.
    • The authors offer a transparent presentation of the tested hyperparameters, which aids reproducibility and guides the reader through the experimental setup in a meaningful way.
    • A variety of experimental setups are employed to rigorously evaluate CytoSAE, supported by in-depth analyses that strengthen the credibility of the findings.
  • 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.
    • While the results convincingly demonstrate the applicability of CytoSAE, the scientific novelty appears limited, as the primary contribution involves applying existing sparse autoencoder (SAE) techniques to histopathology datasets without substantial methodological innovation.
    • Despite a strong introduction, the methodological section lacks sufficient depth and clarity. The rationale behind using a particular MLP architecture to extract meaningful concepts remains underexplained, leaving the underlying intuition unclear.
    • Key variables in equations are not adequately defined, which may hinder comprehension. For instance, in the equations involving z and \hat{x}, the encoder f(x) and decoder g(x) should be explicitly introduced either in the main text or in Figure 1. Similarly, although W and b are commonly used to denote weights and biases, it would be helpful for clarity to define them explicitly.
    • On page 2, the notation d_m is introduced without defining the index m, which could confuse readers. Moreover, while CytoSAE appears to use a multi-layer encoder and decoder, the explanation of how z is derived only describes a single-layer process, lacking completeness.
    • The clustering approach used to identify highly activated latent dimensions is not sufficiently justified and lacks explanation. The motivation for clustering, as well as the choice of k = 10 in the KMeans algorithm, should be better supported—either theoretically or empirically. Additionally, the variable h_{i,j} is introduced without prior explanation, making it difficult to follow.
    • Figure 2a would benefit from a more detailed caption or explanation. It is unclear whether the plotted points represent fixed latent variables, and the underlying assumption—that latents with high activation frequency and strength correspond to meaningful concepts—requires more thorough justification.
    • Finally, in Section 3.4, the reasoning behind selecting layers 7 through 11 for further analysis is not well explained. Clarifying why these layers are presumed to contain the most relevant features would strengthen the argument.
  • 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 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

    I personally enjoyed reading the paper and found the results compelling—CytoSAE appears to be a promising tool for interpretable analysis in hematology. However, to strengthen the impact of the work, I believe the presentation of both the methods and results would benefit from clearer explanations and more thorough rationale behind key design choices. Addressing these points would significantly improve the clarity and make the submission suitable for acceptance in my view.

  • 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.

    (3) Weak Reject — could be rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Although the results presented are convincing and demonstrate the practical effectiveness of the approach, the methodological novelty of the paper is limited. Furthermore, the presentation of the methods, along with the explanations and rationale behind key design choices, is at times unclear and leaves room for interpretation.

  • 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 presents sparse autoencoder for visual concept learning and interpretability applied to hematology images. Five publicly available datasets were used for training and evaluation. Analysis was presented at patch-level, image-level, patient-level, and disease-level.

  • 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.

    Good application area, use of five public datasets for hierarchical analysis

  • 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.

    Results are not presented using any evaluation metric. Showing example images as results does not provide comprehensive information about model performance. For example, patch-wise attribution presented in Fig. 3, what % of Eosinophil cells got the expected activations? Was there any other cell type being false positive? How frequently?

  • 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 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

    Page 4: To identify morphological concepts, we first collected maximally activated reference images for each SAE latent from all five datasets and collaborated with an expert cytomorphologist with approx. 15 years of experience for validation.-> And did what with the reference images? so the maximal activation is based on a_i? How was the expert validation measured? What was the agreement?

    Figure 2: What is meant by label entropy in Figure 2? What is activation freq? Page 5: Expert manual labeling of these concepts (see Sec. 2) confirmed that CytoSAE successfully captured various morphologically relevant features.-> how is the agreement with human expert measured? Please report the values. page 6: Could try similar patch-level analysis using the embeddings from the backbone instead of latents. The distinct pattern of the granules might be found there similarly. Page 7: we identified the top-100 latents that were differentially expressed between the CBFB::MYH11 and PML::RARA subtypes -> Why this two subtypes out of the four?

  • 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?

    Novel Application, Weak evaluation and presentation.

  • 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 paper proposes the first application of sparse auto-encoders (SAE) for haematology data. Specifically, the authors train an SAE on 40K+ single-cell images. The authors propose that as opposed to MILs, SAEs provide a deeper understanding of morphological concepts within an image that contribute to a certain outcome/prediction.

    The disentangled representations obtained using SAE are evaluated against expert annotations, to see if morphological concepts can be extracted from these representations. The method is evaluated on 5 total datasets with labelled single-cell images.

  • 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.

    Evaluation: The evaluation datasets have a large variety of concepts and cover a large subset of general single-cell image data. This is enough proof to show that the model is generalizable to all single-cell images. Similarly, the concept extraction method has potential to act as a benchmark to evaluate single-cell representations for different cell image encoders.

    Results and Ablations: Fig 2A clearly indicates that the concepts are spread across the latent space optimally, as the most occurring concepts have the highest entropy (are most diverse). Fig 2B: visually, each index is showing a certain unique concept across all datasets.
    Adjusting the sparsity regularizer and the expansion factor is well studied, and the effect of each is well explained.

    General clarity: The paper is clear and well explained throughout.

  • 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.

    Classification: As we are working with frozen encoders, the comparisons with other methods that work with frozen encoders (specficially MILs such as ABMIL [1] and their variants) for multi-class classification is not carried out. Ideally, as the representations are disentangled, the information from the barcode should be rich and should lead to strong predictions, however we do not know that as there is no comparison.

    [1] Ilse, Maximilian, Jakub Tomczak, and Max Welling. “Attention-based deep multiple instance learning.” International conference on machine learning. PMLR, 2018.

  • 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 mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure 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

    Downstream tasks: It would be good if the disentangled representations obtained using SAE are shown to outperform MILs that use the input representations on downstream classification tasks.

    Insight into random indexes: It would be good to see what some of the more random dim indexes show. Clearly, the expansion of 64K works best, but this does not mean there are 64K unique concepts, neither does it mean that 16K works better (as shown in the paper). Therefore, I am interested to see what information is carried in those indexes.

  • 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 applies a known method of obtaining disentangled latent representations to Haematology data. Although the application is simple, the thorough experimental validation and presentation of the paper warrants an acceptance.

  • 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

We thank the reviewers for their constructive feedback and recognition of our contributions, including the use of diverse hematology datasets, hierarchical interpretability, and clear visualizations. Below we discuss additions to the camera ready to address the reviewer comments.

R4: Methodological novelty of the paper is limited – CytoSAE provides patch-level attribution [1] and generates attribution maps for morphological features such as eosinophilic granules. Our latent-based method enables hierarchical interpretability from subcellular features to the disease level. It establishes a predefined dictionary for dataset-independent interpretation, addressing limitations of attention-based MIL. To the best of our knowledge, this is the first SAE application to medical imaging with these particular properties.

R1,4: Methodology was not sufficiently outlined – Due to space constraints, we had to focus on CytoSAE’s specific features, referring readers to prior work for architectural details [2]. We would like to clarify the following key concepts: – Reference images: Set of maximally activating images for each SAE latent. Image-level latent activations are computed by summing the number of activating patches (Eq. 3). We then select the top-k images with the highest scores. – Activated Frequency: Fraction of training images for which the latent is activated. A high frequency suggests a common or nonspecific concept. – Mean Activation Value: Average activation strength across activated images. Higher values indicate model confidence. – Label Entropy: Indicates class specificity (low entropy) or diversity (high entropy). Finally, we appreciate R4’ comments regarding notation and will incorporate these suggestions.

R1,3,4: Quantitative evaluation – We will improve the clarity of the quantitative evaluation. We randomly sampled 50 latents with high mean activation (> -3, log scale) and confirmed that all (50/50) captured meaningful patterns based on an expert review. We did not apply the same analysis to low-activation latents, as they mostly reflected background or diffuse signals. Regarding the mentioned latent in Fig. 3 (eosinophilic granules), while these are most common in eosinophils, they can appear in other cells, and not all eosinophils show them prominently. Since latents capture morphology rather than class, defining false positives by cell types might be misleading. Instead, future work could correlate image-level activations with expert-annotated granularity scores for quantitative validation. On disease level quantification, we identified 50 differentially expressed latents between two AML subtypes. Expert annotation confirmed that 32 of these were disease-specific.

R3: Comparison with baselines – We compared to attention-based MIL [3] in Fig. 4. Our approach—logistic regression on patient-level barcodes from SAE latents—achieves comparable performance or outperforms it in certain subtypes [3]. Beyond accuracy, CytoSAE provides interpretable, concept-level representations across scales. We acknowledge that additional comparisons could be conducted in future work (e.g., to Ilse et al.).

R1: Justification of disease choices – AML Hehr [3] includes four genetically defined AML subtypes. While primarily identified via molecular diagnostics, PML::RARA and CBFB::MYH11 show the most distinctive morphological patterns on smears.

R4: Layer selection is not well explained – This is based on testing DinoBloom layers 2, 7, 11, and 12 (Sec. 3.4). Confirming prior work [2], deeper transformer layers yielded SAE latents with richer and more aggregated representations which motivates our choice. When trained on embeddings from shallower layers, the SAE tends to lose meaningful interpretability and captures more generic or low-level features, which are less relevant in our application case.

[1] Le et al., arXiv:2407.10785 [2] Lim et al., ICLR 2025 [3] Hehr et al., PLOS Digit Health 2023




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



back to top