Sparse Autoencoders for Expert-Assisted Interpretation of Histopathology Foundation Model Embeddings

Šimon Šmída; Wanda Benesova

cs.CVOct 29, 2026

Sparse Autoencoders for Expert-Assisted Interpretation of Histopathology Foundation Model Embeddings

Authors: Šimon Šmída, Wanda Benesova

Organizations: Czech Technical University, Prague, Czech Republic

Abstract

Histopathology foundation models learn rich embedding representations, but their internal structure remains difficult to inspect. Concept-level interpretability is particularly relevant in computational pathology because pathologists reason through recurring morphological and tissue-level patterns, including tissue architecture, cellular morphology, staining variation, and artifacts. In this work, we investigate whether sparse autoencoders (SAEs) can decompose embeddings from histopathology foundation models into sparse latent features suitable for expert-assisted interpretation. We train vanilla and Batch-TopK SAE variants on NCT-CRC embeddings, using CONCH as the primary foundation model and UNI2-h as an additional robustness evaluation, and characterize the resulting latent features through reconstruction fidelity, sparsity, coherence, label enrichment, and downstream class-signal preservation. The learned sparse representations preserve most downstream tissue-class information, and their latent features retrieve more coherent and label-concentrated patch groups than raw embedding dimensions or PCA components. Discussions with pathology experts further motivated a WSI-level latent visualization workflow that places feature activations into larger tissue context. Overall, the additional UNI2-h evaluation suggests that the observed behavior is not specific to CONCH, supporting the use of sparse autoencoders for expert-assisted analysis of histopathology foundation-model embeddings, while emphasizing that latent features should be treated as candidate concepts requiring pathological assessment rather than automatically validated histopathological concepts.

Figures & tables

Figure 1: Expert-assisted interpretation of histopathology foundation-model embeddings with sparse autoencoders. (A) Sparse representation learning. Dense patch embeddings are decomposed into sparse latent features using a sparse autoencoder. (B) Expert-assisted latent interpretation. Individual latent features are analyzed through complementary forms of evidence, including quantitative characterization, top-activating patches, and WSI-level activation maps. Together these provide evidence for a candidate concept, which can then be assessed and validated by pathology experts.

Figure 2: Reconstruction–sparsity tradeoff for evaluated SAE variants. BatchTopK SAEs form a consistently better reconstruction–sparsity frontier than vanilla

L_{1}

SAEs across evaluated dictionary widths and sparsity settings. Increasing target sparsity

L_{0}

improves reconstruction fidelity while preserving sparse latent structure.

Figure 3: SAE dictionary utilization. Although the selected BatchTopK SAE reconstructs CONCH embeddings with high fidelity, only a minority of dictionary elements activate on the NCT-CRC dataset. The figure summarizes alive and dead latent features ( left ) together with the activation-frequency distribution of active features ( right ).

Figure 4: Representative SAE latent features as candidate histopathological concepts. Each row shows the highest-activating patches for a selected latent feature together with representative WSI-level deployment patterns. Many latent features exhibit visually coherent and label-enriched activation behavior, although some remain mixed or ambiguous, supporting our interpretation of SAE latents as candidate concepts rather than validated biological abstractions.

Figure 5: SAE latent features retrieve more coherent patch groups than alternative embedding directions. Left: mean pairwise cosine similarity of retrieved patch sets in the original CONCH embedding space. Raw dimensions, PCA components, and random directions are evaluated using the union of positive and negative tails. Right: retrieval examples. SAE latents retrieve coherent patch groups through a single nonnegative activation pattern, whereas raw dimensions can associate different structures with opposite tails.

Figure 6: Sparse latent representations preserve downstream tissue-class information. Left: Linear-probe macro-F1 for full CONCH embeddings, SAE reconstructions, and full SAE latent representations. Most tissue-class information is retained after sparse decomposition. Right: Linear-probe macro-F1 as a function of the number of selected features per tissue class. Selected SAE features retain most classification performance with relatively few features and approach full latent-representation performance as additional features are included. Dashed lines indicate the full-representation baselines.

Backbone	$d$	$m$	NMSE $\downarrow$	EV $\uparrow$	Measured $L_{0}$	Alive (%)
CONCH	512	4,096	$0.01169\!\pm\!0.00014$	$0.97820\!\pm\!0.00027$	$65.72\!\pm\!0.26$	$13.12\!\pm\!0.49$
UNI2-h	1,536	12,288	$0.16733\!\pm\!0.00021$	$0.79792\!\pm\!0.00026$	$223.34\!\pm\!0.48$	$73.28\!\pm\!0.67$

Table 1: Reconstruction and sparse-representation statistics on NCT-CRC for CONCH and UNI2-h foundation models. Values are validation mean

\pm

SD over five training seeds; alive fraction is measured on the validation split.

Backbone	Coherence $\uparrow$	Entropy $\downarrow$	Raw F1	Recon. F1	Latent F1
CONCH	0.8348	0.0898	0.9497	0.9448	0.9358
UNI2-h	0.6052	0.1010	0.9553	0.9573	0.9542

Table 2: Interpretability and downstream evaluation for CONCH and UNI2-h on NCT-CRC. Coherence and entropy summarize top-32 retrieval of SAE features (303 CONCH and 6,118 UNI2-h features), measured in each model’s own embedding space, so absolute values should not be compared across models. F1 values are point estimates from a linear-probe protocol matched across both models (L-BFGS with strong-Wolfe line search); the CONCH values therefore differ slightly from the bootstrap means in Section 4.5 .

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

SAE	Sparsity	$m$	NMSE $\downarrow$	EV $\uparrow$	Cosine $\uparrow$	measured $L_{0}$	Alive %
BatchTopK	$L_{0}=8$	1024	0.1073	0.800	0.946	8.74	10.4
BatchTopK	$L_{0}=64$	1024	0.0141	0.974	0.993	64.00	25.0
BatchTopK	$L_{0}=8$	2048	0.1041	0.806	0.947	9.00	6.3
BatchTopK	$L_{0}=16$	2048	0.0612	0.886	0.969	16.96	9.3
BatchTopK	$L_{0}=32$	2048	0.0259	0.952	0.987	33.60	11.3
BatchTopK	$L_{0}=64$	2048	0.0124	0.977	0.994	65.40	17.3

Appendix

Table 3: Validation metrics for SAE model-selection runs with complete validation outputs. BatchTopK models are specified by target

L_{0}

; vanilla

L_{1}

models are specified by the sparsity coefficient

\lambda

. Alive fraction is measured on the NCT-CRC validation split.

Configuration	NMSE	EV	measured $L_{0}$	Val. alive fraction
$m=4096,\ L_{0}=32$	0.0252 $\pm$ 0.0003	0.9531 $\pm$ 0.0005	33.93 $\pm$ 0.17	0.072 $\pm$ 0.005
$m=4096,\ L_{0}=64$	0.0117 $\pm$ 0.0001	0.9782 $\pm$ 0.0003	65.72 $\pm$ 0.26	0.131 $\pm$ 0.005

Appendix

Table 4: Seed stability for BatchTopK SAEs with

m=4096

. Values are mean

\pm

standard deviation across five random seeds.

$m$	NMSE $\downarrow$	EV $\uparrow$	measured $L_{0}$	active	coherence $\uparrow$	entropy $\downarrow$
1024	0.0141	0.9736	64.00	329 / 1024 (32.1%)	0.833	0.101
2048	0.0124	0.9770	65.40	450 / 2048 (22.0%)	0.828	0.101
4096	0.0115	0.9785	66.03	685 / 4096 (16.7%)	0.835	0.090

Appendix

Table 5: Effect of dictionary width for BatchTopK SAEs at target

L_{0}=64

. Active features are counted on the combined train and validation split. Coherence and entropy are computed from top-32 patch retrieval over eligible active features.

Figure 7: Dictionary-width tradeoff at target $L_{0}=64$ . Increasing dictionary width improves reconstruction fidelity, whereas smaller dictionaries use a larger fraction of their latent features.

Summary over active features	Raw entropy	Normalized entropy	Class-frequency normalized
Median	0.256	0.116	0.122
Mean	0.550	0.250	0.252
25th percentile	0.000	0.000	0.000
75th percentile	0.807	0.367	0.381

Appendix

Table 6: Entropy robustness for active features of the selected SAE. Raw entropy is measured in nats. Normalized entropy divides by

\log 9

. Class-frequency-normalized entropy first reweights each feature’s activation-mass distribution by the inverse train+validation class frequency and then divides by

\log 9

Method	Metric	Top-16	Top-32	Top-64
SAE	Coherence	0.842	0.835	0.822
Raw dimensions	Coherence	0.422	0.427	0.430
PCA	Coherence	0.421	0.417	0.414
Random	Coherence	0.419	0.425	0.429
SAE	Entropy	0.081	0.090	0.103
Raw dimensions	Entropy	0.568	0.596	0.622

Appendix

Table 7: Sensitivity to the number of retrieved top patches. SAE rows use top activations of the selected BatchTopK model. Baseline rows use the union of positive and negative signed-axis tails, as in the main comparison.

Method	Union coherence $\uparrow$	Best-tail coherence $\uparrow$	Union entropy $\downarrow$	Best-tail entropy $\downarrow$
SAE feature	0.835	0.835	0.090	0.090
Raw CONCH dim	0.427	0.749	0.596	0.219
PCA component	0.417	0.473	0.819	0.711
Random direction	0.425	0.748	0.608	0.223

Appendix

Table 8: Sensitivity to signed-tail handling for top-32 retrieval. Raw dimensions, PCA components, and random directions are signed axes. The signed-union setting combines positive and negative tails; the best-tail setting gives each signed axis the more coherent of its two tails. SAE latents are nonnegative, so the same top-activation set is used in both columns.

$k$ per class	SAE/random features	Raw features	SAE selected	Raw selected	Random SAE subset
1	9	9	0.858 $\pm$ 0.003	0.873 $\pm$ 0.002	0.300 $\pm$ 0.111
3	27	27	0.921 $\pm$ 0.002	0.927 $\pm$ 0.002	0.615 $\pm$ 0.084
5	45	45	0.925 $\pm$ 0.001	0.932 $\pm$ 0.002	0.738 $\pm$ 0.064
10	90	89	0.936 $\pm$ 0.001	0.936 $\pm$ 0.002	0.875 $\pm$ 0.029

Appendix

Table 9: Train-only selected-feature probe results for the selected BatchTopK SAE. Features are ranked by absolute Pearson correlation with one-vs-rest class labels on the training split only. The table reports validation macro-F1 mean

\pm

standard deviation. Random subsets sample the same number of alive SAE features as the selected SAE union for each

k

; raw selected dimensions can be fewer when class-wise selections overlap.

Figure 8: Additional coherent and label-enriched latent features. Each panel shows top-activating NCT-CRC patches for one active latent feature, together with compact feature-level statistics from the selected BatchTopK SAE.

Figure 9: Mixed, ambiguous, and artifact-associated latent features. These examples illustrate broad, weakly selective, or technically driven activation behavior and motivate treating SAE latents as candidate concepts requiring expert assessment.

Parameter	Value
WSI source	5 TCGA-COAD SVS slides
Selected SAE	BatchTopK, $m=4096$ , target $L_{0}=64$ , seed 1
Embedding call	CONCH image encoder, unnormalized output
Analysis resolution	$\approx 0.5\,\mu\mathrm{m}$ /px
Encoded patch size	$224\times 224$ pixels
Native WSI footprint / stride	444–456 px / 444–456 px

Appendix

Table 10: TCGA-COAD WSI qualitative overlays. These WSI overlays are used for spatial inspection of candidate concepts.

Figure 10: Screenshot of the WSI-level latent inspection interface. Example view for one selected SAE latent feature on a TCGA-COAD whole-slide image. The interface links top-activating patches with the corresponding slide-level activation overlay, supporting qualitative inspection of spatial deployment within larger tissue context.

Method	Coherence $\uparrow$	Entropy $\downarrow$
SAE feature	0.605	0.101
Raw UNI2-h dim	0.497	0.232
Random direction	0.485	0.249
PCA component	0.290	0.571

Appendix

Table 11: UNI2-h top-32 retrieval compared with signed directions. Each raw dimension (1,536), PCA component (512) and random direction (512) uses its more coherent tail; entropy is measured on the same patch set.

Figure 11: Top-32 retrieval coherence under the best-tail setting for CONCH (left) and UNI2-h (right). For every signed raw, random or PCA direction, the more coherent tail is used; SAE features use their highest positive activations. Each panel is a comparison within one embedding space, so absolute values should not be compared across models.

Figure 12: Candidate features of similar type in independently trained CONCH and UNI2-h SAEs. Top-activating NCT-CRC patches for features associated with lymphocytes, adipose tissue and tumor epithelium. Visual similarity does not establish a one-to-one correspondence between features.

Explainable Medical Image Analysis

Explainable Medical Image Analysis develops methods that make medical image predictions understandable, verifiable, and useful for clinical review.

See all 266 papers

Mar 24, 2026

Sparse Autoencoders for Interpretable Medical Image Representation Learning

Vision foundation models (FMs) achieve state-of-the-art performance in medical imaging. However, they encode information in abstract latent representations that clinicians cannot interrogate or verify. The goal of this study is to investigate Sparse Autoencoders (SAEs) for replacing opaque FM image representations with human-interpretable, sparse features. We train SAEs on embeddings from BiomedParse (biomedical) and DINOv3 (general-purpose) using 909,873 CT and MRI 2D image slices from the TotalSegmentator dataset. We find that learned sparse features: (a) reconstruct original embeddings with high fidelity (R2 up to 0.941) and recover up to 87.8% of downstream performance using only 10 features (99.4% dimensionality reduction), (b) preserve semantic fidelity in image retrieval tasks, (c) correspond to specific concepts that can be expressed in language using large language model (LLM)-based auto-interpretation. (d) bridge clinical language and abstract latent representations in zero-shot language-driven image retrieval. Our work indicates SAEs are a promising pathway towards interpretable, concept-driven medical vision systems. Code repository: https://github.com/pwesp/sail.

Philipp Wesp, Robbie Holland, Vasiliki Sideri-Lampretsa +1

Stanford Center for Artificial Intelligence in Medicine and Imaging, Stanford University, Stanford, CA, USA · Department of Radiology, Stanford University, Stanford, CA, USA · Chair of AI in Healthcare and Medicine, Technical University of Munich, Munich, Germany +1

Medical Image Retrieval Explainable Medical Image Analysis Medical Image Representation Learning Sparse Autoencoders Medical Imaging Foundation Models

Ask PDF

Aug 28, 2026

EXPOSE: Explainable and Domain-Robust Embeddings from Pathology Vision Foundation Models using Sparse Autoencoders

Vision Foundation Models (VFMs) are widely used in computational pathology but remain sensitive to domain shifts arising from variations in staining, tissue preparation, and scanner hardware. A key limitation is that VFM embeddings entangle biological with domain-specific information, hindering cross-domain generalization. We propose Explainable Probing of Cross-Domain Sparse Embeddings (EXPOSE), a framework that uses Sparse Autoencoders (SAEs) as an explainable bottleneck to identify and suppress domain-specific components in VFM embeddings. We train a sparse representation of VFM features, use a linear classifier to identify domain-specific latent dimensions, and mask these features prior to downstream relapse prediction without retraining the backbone model. Experiments on a large prostate cancer dataset with multiple acquisition domains show that SAE features capture both domain- and task-specific information, which are partially disentangled in the latent space. Removing domain-specific features improves cross-domain performance and increases embedding robustness as measured by the Domain Robustness Index (DoRI). Code is available at https://github.com/imsb-uke/expose .

Anja Witte, Maximilian Lennartz, Jan Baumbach +4

Institute of Medical Systems Bioinformatics, Center for Biomedical AI (bAIome), Center for Molecular Neurobiology Hamburg (ZMNH), Hamburg Center for Translational Immunology (HCTI), University Medical Center Hamburg-Eppendorf, Hamburg, Germany · Institute of Pathology, University Medical Center Hamburg-Eppendorf, Hamburg, Germany · Institute of Computational Systems Biology, University of Hamburg, Hamburg, Germany +2

Computational Pathology Pathology Foundation Models Explainable Medical Image Analysis Sparse Autoencoders

Ask PDF

Apr 13, 2026

Sparse Autoencoder Decomposition of Clinical Sequence Model Representations: Feature Complexity, Task Specialisation, and Mortality Prediction

Sparse autoencoders (SAEs) have been applied to large language models and protein language models, but not systematically to electronic health record (EHR) foundation models. We train TopK SAEs on FlatASCEND, a 14.5-million-parameter autoregressive clinical sequence model, at all 10 residual stream extraction points on INSPECT (outpatient) and MIMIC-IV (ICU). SAE decomposition reveals progressive abstraction across transformer depth: layer-0 features are near-perfect token detectors (45.7% singleton), while layer-6 features span approximately 30 token types across multiple clinical categories (0.5% singleton). Under full-sequence simple linear probes, SAE features outperform dense representations for discrete event prediction (mortality) while dense representations outperform for continuous magnitude prediction (length of stay) - a probe-level representational phenomenon that does not extend to clinically relevant leakage-safe windows, where dense representations match or exceed SAE features across all tested settings (eICU-CRD 48-hour AUC: SAE 0.871 versus dense 0.880; base model zero-shot, SAE dictionaries trained on eICU activations; MIMIC-IV: 0.836 versus 0.914; INSPECT 1-year/3-year: 0.697 versus 0.800). A delta-mode intervention method reduces SAE perturbation noise by 86x, enabling cleaner feature-level experiments, though the resulting perturbation effects are larger than random controls in 3 of 4 conditions but not formally significant. Feature reproducibility across random seeds is 21%, and individual features should be interpreted as illustrative rather than stable.

Chris Sainsbury, Feng Dong, Andreas Karwath

School of Cardiovascular & Metabolic Health, University of Glasgow, UK · School of Medicine, University of Dundee, UK · NHS Greater Glasgow and Clyde, UK +2

EHR Foundation Models Sparse Autoencoders Mechanistic Interpretability Electronic Health Records Clinical Outcome Prediction Longitudinal EHR Modeling

Ask PDF

Abstract

Explainable Medical Image Analysis

Explore similar work

Sparse Autoencoders for Interpretable Medical Image Representation Learning

EXPOSE: Explainable and Domain-Robust Embeddings from Pathology Vision Foundation Models using Sparse Autoencoders

Sparse Autoencoder Decomposition of Clinical Sequence Model Representations: Feature Complexity, Task Specialisation, and Mortality Prediction