{"id":"cf7f3917-b965-4e9d-8062-a6016c1b5294","arxiv_id":"2507.12464","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CytoSAE learns sparse, expert-validated morphological concepts from blood-cell images that generalize across datasets and can classify AML subtypes at patient level with F1 0.83.","lead":"Researchers trained a sparse autoencoder on embeddings from a blood-cell foundation model to break cell images into interpretable visual concepts. The method transfers to bone marrow samples and can summarize a patient's AML subtype as a count of these concepts, with accuracy close to a black-box deep learning model.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"AML classification result is not yet attributable to CytoSAE concepts: no patient-stratified split details, no raw-embedding linear-probe baseline, and threshold selection on the same data.","rationale":"The reader's weakest assumption concerns whether DinoBloom-B residual-stream features are stable, linearly decodable morphology rather than style/batch artifacts. That is important, but even if the backbone features are perfectly morphological, the paper's central quantitative demonstration of CytoSAE's value is the AML classification result. The missing raw-embedding baseline and unstated split protocol directly undermine this demonstration, and threshold selection on the same data adds optimistic bias. I chose partial agreement because the classification-protocol gap is the most load-bearing concern I can identify, while the reader emphasizes the backbone-stability assumption. Both are addressable with a controlled re-run, so the conditional verdict remains appropriate rather than escalating to reject.","tokens_in":7645,"tokens_out":5875,"duration_ms":72295,"concrete_test":"Re-run the AML subtype task with patient-stratified 5-fold cross-validation. On each training fold, select the threshold (or number of latents) using only that fold. Report mean±std weighted F1 for: (a) CytoSAE barcodes; (b) a linear probe on raw DinoBloom-B features aggregated per patient (e.g., mean of CLS or mean-pooled patch tokens); (c) barcodes from an untrained/random SAE as a null model. If (a) is not significantly above (b) and (c), the F1 does not demonstrate CytoSAE-specific interpretable predictive power; if (a) drops sharply under patient-stratified splitting, leakage explains the reported number.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing quantitative claim is the weighted F1 of 0.832±0.044 for AML subtype classification from CytoSAE patient barcodes (Sec. 3.3, Fig. 4B). For this number to establish that SAE concepts carry disease-relevant morphology, the evaluation must prevent information leakage and compare against embeddings that the SAE merely compresses. The paper does not state whether classification used patient-stratified k-fold cross-validation or a fixed split, does not report the number of folds or runs behind the ±0.044 interval, and gives no linear-probe baseline on raw DinoBloom-B embeddings (e.g., patient-level mean of CLS or patch tokens). It also selects the threshold -3 by varying the number of latents and observing performance on the same data, so the reported F1 may be optimistically biased by threshold selection, especially with 189 patients and roughly 2,936 features. Without these controls, the comparable-to-SOTA claim could be an artifact of leakage or of the underlying DinoBloom features rather than of CytoSAE's discovered morphology.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CytoSAE, a sparse autoencoder trained on DinoBloom-B residual-stream tokens extracted from 41,906 peripheral blood single-cell images. The model expands 768-dimensional tokens to a 49,152-dimensional sparse latent space using an MSE reconstruction loss with an L1 penalty, then binarizes patch-level activations and aggregates them into image-, patient-, and disease-level barcodes. The authors report that the discovered latents correspond to morphologically meaningful concepts, that these concepts transfer across four external datasets including bone marrow cytology, and that logistic regression on patient barcodes classifies four AML subtypes plus a healthy control group with weighted F1 0.832±0.044, comparable to a fully deep learning baseline, while offering sub-cellular patch-level explanations.","tokens_in":7896,"tokens_out":4864,"duration_ms":55481,"significance":"If the claims hold, CytoSAE is a valuable interpretability layer for hematology foundation models: it provides a structured, patch-attributable vocabulary of morphological concepts and a transparent patient-level representation for subtype classification. The paper's strengths include the release of source code and model weights, the use of established public datasets, the involvement of an expert cytomorphologist, and explicit ablation studies over expansion factors, sparsity coefficients, residual layers, and ghost gradient resampling. The main caveat is that the quantitative claims, especially the classification result and the disease-specific concept validation, need stronger evaluation protocols before the conclusions can be fully supported.","major_comments":[{"comment":"The key quantitative claim, weighted F1 0.832±0.044 for AML subtype classification from patient barcodes, lacks essential evaluation detail. The paper does not state whether classification used patient-stratified k-fold cross-validation or a fixed split, how many folds or runs produced the ±0.044 interval, or how the 189 patients were partitioned. Because barcodes aggregate multiple images per patient, any non-patient-stratified split can leak information between training and test sets. In addition, no linear-probe baseline on raw DinoBloom-B embeddings is reported; without such a baseline on, for example, patient-level mean CLS or patch tokens under the identical protocol, the result cannot be attributed to CytoSAE concepts rather than to the backbone features. Please report the exact evaluation protocol and add this baseline.","section":"Sec. 3.3, Eq. (3), Fig. 4"},{"comment":"The activation threshold tau in Eq. (2) is selected by varying the number of latents and monitoring classification performance on the same data. With 189 patients and roughly 2,936 features at the chosen threshold, selecting the threshold on the evaluation data can optimistically bias the reported F1. The threshold selection procedure should be described on a validation split or through nested cross-validation, and performance should be reported across the full threshold range with error bars, not only at the selected threshold of -3.","section":"Sec. 3.3 and Fig. 4A"},{"comment":"The expert validation of disease-specific concepts is ambiguous and weak for PML::RARA. The text reports \"5 out of 10 and 32 out of 50\" for CBFB::MYH11-specific concepts and \"4 out of 10 and 10 out of 50\" for PML::RARA-specific concepts, but the sample sizes and the relationship between the two numbers are unclear. As written, only 10 out of 50 of the top PML::RARA features were judged disease-specific, which substantially weakens the claim that disease-specific concepts are discovered. Please clarify the full contingency table, report how the 10 and 50 subsets were constructed, and compare against chance agreement.","section":"Sec. 3.3, disease-barcode paragraph"},{"comment":"The cross-dataset generalization claim is supported mainly by qualitative visualization of selected latents and images. No quantitative measure of concept consistency across datasets is provided, so it is difficult to distinguish shared morphology from backbone style or batch effects. A quantitative evaluation, such as measuring how often the top activating patches for a latent in MLL23 are also top activations in Acevedo, Matek19, or BMC, or comparing activation statistics across datasets, would make the transfer claim more robust.","section":"Sec. 3.1, Fig. 2"}],"minor_comments":[{"comment":"The caption reads \"within the patent's single-cell images\" and should read \"within the patient's single-cell images.\"","section":"Fig. 1B caption"},{"comment":"There are typos: \"influced\" should be \"influenced\" and \"treshold\" should be \"threshold.\"","section":"Sec. 3.4"},{"comment":"The threshold tau is used in Eq. (2) but never explicitly defined; the text later refers to thresholds in log10 mean activation units. Please define tau in the text or state that it is the mean activation threshold.","section":"Sec. 2, Eq. (2)"},{"comment":"The comparison to the deep learning baseline [12] is reported only as \"similar\" without a statistical comparison; please provide confidence intervals or a significance test for the difference in weighted F1.","section":"Sec. 3.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and the proposed method is promising. The main risk is that the central quantitative claims are not yet backed by a clearly described evaluation protocol, and the disease-specific concept validation is weak. These issues are addressable with additional experiments and clearer reporting, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"CytoSAE is the first sparse autoencoder applied to hematology cell images, and the qualitative core holds up. The authors train a 49,152-unit SAE on DinoBloom-B residual tokens from ~42k peripheral blood cells, then show the discovered concepts transfer to four other datasets including bone marrow, with patch-level attribution that is visually convincing. The patient- and disease-level barcode aggregation is a new analytical step for this domain, and they ship code and weights, which makes the whole thing reproducible. The eosinophil granule example is a clean demonstration of sub-cellular attribution.\n\nThe soft spots are real but not disqualifying. Expert validation is thin: 50 latents sampled via k-means, and in the disease-specific comparison only 10 of 50 PML::RARA concepts were judged meaningful. That is reported honestly, but it does not support a strong claim about disease-specific morphology for that subtype. The AML classification result (weighted F1 0.832±0.044) is not yet attributable to CytoSAE. There is no patient-stratified split description, no raw-embedding linear-probe baseline, and the activation threshold -3 was chosen by varying the number of latents and observing performance on the same data. With 189 patients and ~2,936 features, selection bias and possible leakage are real concerns. The stress-test note is on target; I don't think the authors cheated, but the F1 needs controls before it can be compared to the deep-learning baseline. Additionally, the claim that performance remains stable until meaningful concepts are removed is confounded by threshold selection on the same data.\n\nThe central qualitative claim—that a SAE on residual-stream features discovers stable, transferable morphological concepts—holds up. This is a legitimate application of an established method to a new domain, with a thoughtful evaluation structure and honest limitation statements. The citation pattern is fine; the backbone and training data come from the same group, but that is disclosed and reasonable for a methods-application paper. It deserves peer review. The referee should ask for: (1) split/fold details for the barcode classifier and a raw DinoBloom-B embedding baseline; (2) held-out threshold selection or nested validation; (3) more quantitative expert validation, particularly for PML::RARA. Those are fixable.\n\nWho is this for? Anyone working on interpretability for medical imaging, especially hematology foundation models. It is not a foundational methods paper, but it is a solid domain application with reproducible artifacts. Recommendation: accept for peer review with the expectation of a revision.","headline":"A genuinely new application of SAEs to hematology with a credible qualitative core, but the AML classification F1 needs baseline and split controls before it carries weight.","tokens_in":8407,"tokens_out":3394,"would_cite":false,"duration_ms":35752,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CytoSAE shows that sparse autoencoders can turn hematology embeddings into transferable morphological concepts and explainable AML-subtype classifiers.","keywords":["sparse autoencoder","hematology","cell morphology","interpretability","acute myeloid leukemia","concept discovery","foundation models","patch-level attribution"],"falsifier":"One decisive test would be to take a latent that fires on eosinophilic granules in peripheral blood and check whether the same latent index fires on eosinophilic granules in bone marrow images from a different institution, scanner, and stain protocol; if the mapping breaks, the cross-dataset consistency is an acquisition artifact rather than shared morphology. A second check is to train the identical SAE on a re-initialized DinoBloom-B backbone and see whether the same concepts reappear.","tokens_in":7501,"feed_emoji":"🩸","tokens_out":13767,"duration_ms":138709,"temperature":0.7,"pith_summary":"This paper sets out to show that sparse autoencoders, a dictionary-learning tool from interpretability research, can make a general-purpose cell-image foundation model explainable at the sub-cellular level. The authors train CytoSAE on the internal per-patch vectors of DinoBloom-B, a hematology foundation model, using 41,906 peripheral-blood single-cell images and expanding each 768-dimensional vector into 49,152 sparse latent units. They argue that the learned units are genuine morphological concepts because the same units fire on matching structures across four external datasets, including bone marrow cytology, and because an expert cytomorphologist confirms the recovered patterns. Counting active units per patient produces barcodes on which logistic regression reaches a weighted F1 of $0.832 \\pm 0.044$ for four genetic AML subtypes and healthy controls, comparable to the full deep-learning baseline. The contribution is a concrete route from black-box cell embeddings to disease-specific, visually checkable evidence.","feed_headline":"Sparse autoencoder turns blood-cell images into explainable concepts","feed_subtitle":"Trained on 41,906 peripheral-blood images, the model's concepts transfer to bone marrow and pinpoint AML subtype features.","key_machinery":"The carrying mechanism is a sparse autoencoder with a 64-fold expansion: a 768-dimensional token $x$ from the backbone's residual stream is encoded as $z = \\mathrm{ReLU}(W_{\\mathrm{enc}}(x - b_{\\mathrm{dec}}) + b_{\\mathrm{enc}}) \\in \\mathbb{R}^{49{,}152}$ and decoded as $\\hat{x} = W_{\\mathrm{dec}} z + b_{\\mathrm{dec}}$, trained by MSE reconstruction plus an L1 penalty on $z$, with the decoder bias initialized at the geometric median of training tokens and ghost-gradient resampling to revive dead latents. Each latent index $s$ is treated as one morphological concept. A threshold $\\tau$ binarizes patch-level activations, counting active patches to form image-, patient-, and disease-level barcodes; multiplying a patch by its activation yields a concept-attribution map. This hierarchical counting is what converts an unsupervised dictionary into a patient-level diagnostic feature space.","core_discovery":"On the paper's own terms, the central claim is that CytoSAE disentangles the compact embedding of a hematology foundation model into a reusable dictionary of morphological concepts, and that these concepts carry the diagnostic signal. The model decomposes the layer-11 residual stream of DinoBloom-B with an expansion factor of 64, producing 49,152 latents; at a mean-activation threshold of $-3$ on a log scale, roughly 2,936 concepts are confidently activated. The same latents generalize to peripheral blood datasets and to 171,373 expert-annotated bone marrow cells, and patch-level activation maps localize structures such as eosinophilic granules consistently across datasets. When patient-level barcodes are classified, the weighted F1 of $0.832 \\pm 0.044$ is comparable to the deep-learning baseline, and differential disease barcodes identify subtype-specific morphologies such as vacuolated monocytic precursors for CBFB::MYH11 and granulated blasts for PML::RARA.","pith_inferences":["Beyond the paper, if the cross-dataset consistency is biological rather than acquisition-driven, the same dictionary could serve as a fixed morphological atlas for rare hematologic diseases, but the paper does not test rare-disease coverage.","A natural extension would be applying the same SAE recipe to residual streams of other medical foundation models; the paper only tests one backbone.","Because only about 2,936 of 49,152 latents are confidently activated, most of the dictionary capacity is idle; pruning to the active subset could make the method cheaper at deployment, though the paper does not evaluate this.","The barcode representation is linear in concept counts, so it could be plugged into survival or regression models with clinical covariates, while the paper demonstrates classification only."],"forward_implications":["A single SAE trained on peripheral blood can be applied directly to bone marrow cytology, so explainable concept extraction does not require retraining per specimen type.","Patch-level attribution makes the basis of a patient-level prediction inspectable, letting clinicians verify that the model is looking at structures such as eosinophilic granules rather than at global slide artifacts.","Keeping only confidently activated concepts improves classification stability, which suggests that pruning the dictionary to meaningful latents is safe for downstream tasks.","Disease-level barcodes provide a data-driven shortlist of morphological hallmarks for a genetic AML subtype, which can then be checked by expert review."],"supporting_citations":[{"why":"Supplies the DinoBloom-B backbone whose layer-11 residual-stream tokens are the input to the SAE, and the logistic-regression setup for classification.","marker":"[14]"},{"why":"Training data source: the 41,906 peripheral-blood single-cell images with 18 cell types used to fit CytoSAE.","marker":"[25]"},{"why":"Provides the sparse autoencoder methodology, including L1 reconstruction, geometric-median decoder bias, and ghost-gradient resampling.","marker":"[4]"},{"why":"Establishes visual-domain SAE concept discovery and the patch-count aggregation that CytoSAE adapts to hematology.","marker":"[17]"},{"why":"Supplies the AML patient dataset with four genetic subtypes plus healthy controls, and the deep-learning baseline whose weighted F1 is compared against.","marker":"[12]"},{"why":"External peripheral-blood dataset used to test whether the discovered concepts transfer to images from a different source.","marker":"[3]"},{"why":"Second external peripheral-blood dataset with expert labels used in concept validation.","marker":"[21]"},{"why":"Bone marrow cytology dataset that tests generalization across specimen types.","marker":"[20]"}],"fun_headline_variants":["Sparse autoencoder decodes blood-cell images into interpretable concepts","CytoSAE maps blood-cell images to explainable concepts","Sparse autoencoder reveals disease-specific concepts in blood smears","Sparse autoencoder pinpoints AML subtype features in blood cells","Sparse autoencoder concepts transfer from blood to bone marrow"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"CytoSAE is only a linear dictionary on top of DinoBloom-B's layer-11 residual stream, so everything depends on those embeddings already storing cell morphology in a stable, sparse, linearly decodable form; if they mostly store staining or scanner style, the cross-dataset consistency would be an artifact of shared acquisition rather than shared biology.","fun_headline_variants_meta":{"raw":{"variants":["Sparse autoencoder decodes blood-cell images into interpretable concepts","CytoSAE maps blood-cell images to explainable concepts","Sparse autoencoder reveals disease-specific concepts in blood smears","Sparse autoencoder pinpoints AML subtype features in blood cells","Sparse autoencoder concepts transfer from blood to bone marrow"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002208,"raw_usage":{"total_tokens":8565,"prompt_tokens":979,"completion_tokens":7586,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":7500}},"tokens_in":595,"tokens_out":7586,"duration_ms":61400,"temperature":1.0,"reasoning_tokens":7500,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:44:49.884138+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One decisive test would be to take a latent that fires on eosinophilic granules in peripheral blood and check whether the same latent index fires on eosinophilic granules in bone marrow images from a different institution, scanner, and stain protocol; if the mapping breaks, the cross-dataset consistency is an acquisition artifact rather than shared morphology. A second check is to train the identical SAE on a re-initialized DinoBloom-B backbone and see whether the same concepts reappear.","supporting_citations":[{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"Supplies the DinoBloom-B backbone whose layer-11 residual-stream tokens are the input to the SAE, and the logistic-regression setup for classification."},{"cited_title":"medRxiv pp","cited_arxiv_id":null,"evidence_quote":"Training data source: the 41,906 peripheral-blood single-cell images with 18 cell types used to fit CytoSAE."},{"cited_title":"PLOS Digital Health2(3), e0000187 (2023)","cited_arxiv_id":null,"evidence_quote":"Supplies the AML patient dataset with four genetic subtypes plus healthy controls, and the deep-learning baseline whose weighted F1 is compared against."},{"cited_title":"https: //doi.org/10.7937/tcia.2019.36f5o9ld, data set","cited_arxiv_id":null,"evidence_quote":"Second external peripheral-blood dataset with expert labels used in concept validation."},{"cited_title":"https://doi.org/10.7937/TCIA.AXH3-T579, data set","cited_arxiv_id":null,"evidence_quote":"Bone marrow cytology dataset that tests generalization across specimen types."}],"review_version":1}