Summary
Understanding the complex patterns inside AI models that analyze tissue images is hard. The authors explore how a special type of neural network called a sparse autoencoder can break down these complex model outputs into simpler parts that experts can interpret. Their method keeps important tissue information while grouping similar tissue patterns together more clearly than other techniques. Pathologists found this approach helpful for visualizing tissue features in a wider context, but the features still need expert review before being considered definitive. This method works across different foundation models, suggesting it could be broadly useful for pathology AI analysis.
What this means in practice
- •For computational pathology teams: Decompose complex pathology image model outputs into simpler features that pathologists can interpret to assist diagnosis workflows.
- •For medical image analysis tool developers: Incorporate sparse autoencoders to generate interpretable tissue feature representations for improved downstream classification and visualization.
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.