CytoSAE: Sparse Autoencoder for Hematology

• CytoSAE is a specialized sparse autoencoder framework for hematological image analysis that generates interpretable, concept-level embeddings.
• It employs a two-layer encoder-decoder architecture with L1 sparsity to disentangle distinct sub-cellular morphological features.
• Validated on diverse datasets, CytoSAE supports reliable patient- and disease-level explainability for AI diagnostic systems.

CytoSAE is a specialized sparse autoencoder framework developed to provide interpretable, concept-level embeddings for hematological single-cell image analysis. Designed for the elucidation of morphological concepts within peripheral blood and bone marrow cytology images, CytoSAE enables sub-cellular, patient-level, and disease-level explainability of AI-based diagnostic systems in hematology. Its architecture and methodology are tailored to extract tangible, clinically-validated features from high-dimensional representations generated by state-of-the-art foundation models in medical imaging, offering both classification performance and mechanistic insight (Dasdelen et al., 16 Jul 2025).

1. Architectural Foundations and Training

CytoSAE is formulated as a two-layer neural network comprising a linear encoder, a non-linear activation, and a linear decoder. The framework operates on image patch tokens generated by DinoBloom-B, a foundation model for hematology. The encoding process is a linear transformation followed by ReLU non-linearity, with reconstruction performed via another linear transformation. The model enforces sparsity on the latent representations through an L1 penalty, supporting disentanglement of distinct morphological factors.

The encoder and decoder are mathematically defined as

$$z = f(x) = \mathrm{ReLU}(W_{\text{enc}} \cdot (x - b_{\text{dec}}) + b_{\text{enc}})$$

$$\hat{x} = g(z) = W_{\text{dec}} \cdot z + b_{\text{dec}}$$

where $x \in \mathbb{R}^{d_m}$ is an input token, $$W_{\text{enc}} \in \mathbb{R}^{d_{\text{SAE}} \times d_m}$$, $$W_{\text{dec}} \in \mathbb{R}^{d_m \times d_{\text{SAE}}}$$, and $b_{\text{enc}}, b_{\text{dec}}$ are bias vectors.

The training objective is given by

$$\min_\theta \sum_{i=1}^K \left\|x_i - \hat{x}_i\right\|_2^2 + \lambda \left\|f(x_i)\right\|_1$$

where $\lambda$ modulates the sparsity constraint.

Training includes additional mechanisms such as ghost gradient resampling to mitigate the presence of inactive latent features, and systematic hyperparameter optimization (e.g., expansion factor, learning rate, decoder bias initialization). The approach targets both minimal reconstruction error and maximal sparsity to ensure interpretability of the latent space (Dasdelen et al., 16 Jul 2025).

2. Application to Hematology

CytoSAE is trained on tokens from more than 40,000 peripheral blood single-cell images in the MLL23 dataset. DinoBloom-B processes individual cell images into a sequence of patch-level tokens; these are then mapped by CytoSAE into a sparse latent space. The outcome is a high-dimensional embedding space in which each latent component is driven to encode a distinct morphological pattern (e.g., specifiable nuclear features, cytoplasmic texture, or granulation characteristics).

To analyze samples at various levels:

• Patch-level: Each image patch’s activation in the latent space indicates the presence of a localized morphological concept.
• Image-level: Patch activations are aggregated to obtain an image “barcode.”
• Patient-level: Aggregation across images from a single patient yields a patient-specific barcode.
• Disease-level: Patient barcodes are further aggregated by disease group for disease-specific interpretation.

Generalizability is established by validating CytoSAE on out-of-domain datasets (including Acevedo, Matek19, BMC, and AML Hehr), encompassing both peripheral blood and bone marrow images. The persistence of concept-level structure across domains demonstrates CytoSAE’s robustness to data heterogeneity, a frequent challenge in clinical imaging (Dasdelen et al., 16 Jul 2025).

3. Concept Extraction and Expert Validation

After training, maximally activated patch tokens for each latent dimension are identified across the validation datasets. Binarization of activations (with a fixed threshold $\tau$) is performed to generate concept barcodes: $a_{(i,j)}[s] = \mathbb{I}(h_{(i,j)}[s] > \tau)$

$a_i[s] = \sum_{j=1}^{n_i} a_{(i,j)}[s]$

$a_p[s] = \frac{1}{N_p} \sum_{i \in I_p} a_i[s]$

$a_d[s] = \frac{1}{N_d} \sum_{p \in P_d} a_p[s]$

where $a_{(i,j)}[s]$ is the binary activation for latent $s$ in patch $j$ of image $i$, and $a_p$, $a_d$ are aggregate activations at the patient and disease levels, respectively.

For comprehensive validation, high-activation patch samples for each latent are clustered (via KMeans with $k=10$) to efficiently sample the diversity of discovered concepts. These are then reviewed by a cytomorphologist with approximately 15 years of experience, who annotates whether each latent corresponds to a morphologically meaningful cellular feature. This expert-driven procedure is critical for ensuring that CytoSAE’s learned representations are directly interpretable and clinically relevant (Dasdelen et al., 16 Jul 2025).

4. Classification, Explainability, and Performance

CytoSAE’s aggregation mechanism is used to generate interpretable feature vectors (“barcodes”) that serve as input for downstream classification. In a demonstrative task on the AML Hehr dataset, a logistic regression classifier trained on patient barcodes achieved a weighted F1-score of approximately 0.832 ± 0.044, a performance level comparable to earlier deep-learning-only approaches. Investigations into the impact of the activation threshold and the number of selected latents indicated that a subset of stable, meaningful concepts is essential for consistent classification accuracy.

Explainability is achieved through patch-wise attribution: for any image, latent activations can be mapped back to individual patches, and those patches can be visualized for their contribution to specific concepts. This enables both visual comparability and interpretability at the sub-cellular level. By weighting image patches according to their latent activation, CytoSAE produces segmentation maps that highlight relevant morphological structures, supporting both algorithmic transparency and clinical inspection (Dasdelen et al., 16 Jul 2025).

5. Implications for Mechanistic and Clinical Understanding

CytoSAE facilitates discovery and annotation of disease- and patient-specific morphological patterns at multiple hierarchies (patch, image, patient, disease). This enables:

• Identification of pathognomonic cells and localized abnormalities, offering mechanistic insight into disease manifestation.
• Direct linkage of deep learning representations to tangible features (e.g., nuclear segmentation, cytoplasmic granularity), providing a foundation for clinical interpretation and potential scientific discovery.
• Cross-dataset and cross-domain generalization, allowing application of CytoSAE-learned concepts to diverse clinical imaging settings.

A crucial consequence is the potential for improved transparency and trust in AI-driven hematology diagnostics, addressing a key limitation of opaque, high-performance neural models in clinical workflows (Dasdelen et al., 16 Jul 2025).

6. Limitations and Future Directions

Whilst CytoSAE demonstrates robustness and interpretability, several extensions are outlined:

• Introduction of patient-specific retrieval algorithms to further personalize diagnoses, with attention to minimizing selection biases.
• Integration of semi-supervised learning to refine and stabilize concept discovery by leveraging both labeled and unlabeled data.
• Expansion to additional datasets and hematological conditions to assess clinical breadth and generalizability.
• Continuous refinement of expert validation protocols to adapt to evolving concept definitions and ensure persistent clinical relevance.

These suggested directions target the systematic enhancement of CytoSAE’s clinical integration, interpretability, and scientific value (Dasdelen et al., 16 Jul 2025).

---

In summary, CytoSAE provides a scalable, interpretable framework for decomposing hematology foundation model representations into medically meaningful morphological features. Its integration of sparse autoencoding, multi-level barcode aggregation, and expert-validated concept schema advances the explainability, performance, and practical utility of AI in hematological image analysis.