CytoSAE

CytoSAE is a sparse autoencoder that makes the embeddings of a hematology imaging foundation model interpretable. Foundation models for single-cell blood and bone marrow images produce dense feature vectors that support strong downstream classification, but those vectors offer no account of why a cell was represented the way it was — a serious obstacle in diagnostic hematology, where a cytomorphologist's judgment rests on nameable features such as granulation, nuclear shape, or cytoplasmic texture. CytoSAE decomposes those dense embeddings into a large dictionary of sparsely activating latents, each of which tends to correspond to a single morphological concept that can be located on the image at the patch level.

The model was developed by researchers at Helmholtz Munich (Institute of AI for Health and Institute of Computational Biology), the Technical University of Munich's hematology and oncology department, and KAIST, and presented at MICCAI 2025. It adapts the PatchSAE framework — originally demonstrated on natural-image vision transformers — to medical cytology, and is trained on frozen DinoBloom-B embeddings rather than on pixels, so the backbone is never modified and the interpretability layer can be attached post hoc.

It is the imaging counterpart to PLM-SAE, which applies the same sparse-dictionary idea to protein language models, and it works on the same blood smear cytology data as GenBloom and Uni-Hema, but with explanation rather than diagnostic accuracy as its objective.

#Key Features

• Concept dictionary over a frozen backbone: 49,152 sparse latents are learned on top of unmodified DinoBloom-B features, so no retraining of the underlying foundation model is required.
• Sub-cellular attribution: Latent activations are computed per image patch, producing segmentation-like maps that localize a concept — eosinophilic granules, for instance — to specific regions within a single cell.
• Cross-dataset and out-of-domain transfer: A single SAE trained on peripheral blood generalizes to independent blood datasets and to bone marrow cytology without refitting, with concepts remaining consistent under domain shift.
• Expert-validated concepts: A cytomorphologist with roughly 15 years of experience annotated sampled latents, confirming that many correspond to genuine, disease-specific morphology.
• Patient-level concept barcodes: Aggregating latent activations across a patient's cells yields a compact, interpretable feature vector that supports AML subtype classification.
• Released code and weights: The implementation is MIT-licensed, and the trained checkpoint is deposited under CC BY 4.0 in Helmholtz Munich's public repository.

#Technical Details

CytoSAE is trained on the 768-dimensional residual stream output of the second-to-last attention block of DinoBloom-B, a DINOv2 ViT-B/14 backbone for hematology. The autoencoder uses an expansion factor of 64, giving 49,152 latents, with sparsity induced by an L1 penalty with coefficient 8×10⁻⁵; expansion factors of 16, 32, 64, and 128 were compared before settling on 64. Training used the MLL23 dataset of 41,906 peripheral blood single-cell images spanning 18 cell types. Evaluation covered Acevedo (17,092 images, 11 classes), Matek19 (18,365 expert-labeled images, 15 classes), the BMC bone marrow set (171,373 annotated cells), and the AML Hehr cohort of 189 patients spanning four genetically defined AML subtypes plus healthy controls.

For the downstream task, patch-level activations are binarized and summed into per-patient "barcodes" that are fed to an L2-regularized logistic regression. At an activation threshold of 10⁻³, five-class discrimination among PML::RARA, NPM1-mutated, CBFB::MYH11, RUNX1::RUNX1T1, and healthy controls reached a weighted F1 score of 0.832 ± 0.044, similar to a fully deep-learning baseline while remaining traceable to individual morphological concepts. In the expert review, 50 latents sampled by k-means clustering (k = 10, five per cluster) were annotated; 32 of the 50 were judged CBFB::MYH11-specific and 10 of the 50 PML::RARA-specific.

#Applications

CytoSAE is aimed at researchers and clinical scientists building or auditing AI systems for blood and bone marrow cytology. It lets a team inspect what a hematology foundation model has actually learned, surface morphologies associated with a particular genetic AML subtype, and present a classification decision alongside the sub-cellular evidence that drove it — a prerequisite for regulatory scrutiny and clinical trust. Because it attaches to frozen embeddings, it is also a general recipe for adding interpretability to other medical imaging foundation models without retraining them.

#Impact

The work extends mechanistic interpretability from language models and natural-image vision transformers into diagnostic medical imaging, a domain where explanation carries clinical weight rather than only scientific interest. Peer-reviewed at MICCAI 2025 and released with both code and weights, it provides a reusable template for concept extraction from frozen medical vision backbones. Its limits are worth stating plainly: interpretability is inherited from and bounded by the DinoBloom-B backbone, expert validation was performed at the dataset level rather than per patient in order to avoid selection bias, and the demonstrated out-of-domain generalization to bone marrow cytology is qualitative, established through expert-reviewed concept consistency rather than a quantitative transfer benchmark.