Explaining Digital Pathology Models via Clustering Activations

A new explainability method for digital pathology models moves beyond single-slide saliency maps to reveal global model behavior through clustering, according to research posted on arXiv. The technique was evaluated on a prostate cancer detection model to demonstrate its clinical utility [1]. The approach, authored by Vít Musil, targets convolutional neural networks used in digital pathology. Unlike occlusion, GradCAM, or relevance propagation methods that highlight influential regions on individual slides, this technique clusters activations to show how the model behaves across many samples [1][2]. The resulting clusters can be visualized to understand the model and increase confidence in its operation, which the paper argues will lead to faster adoption in clinical practice [1][2]. Convolutional neural networks are a class of deep learning architecture commonly applied to medical image analysis [4]. Deep learning models use multiple processing layers to learn representations of data, and have produced results comparable to human expert performance in some medical imaging tasks [4]. However, their complexity often makes them opaque to clinicians, fueling demand for interpretability tools. The paper evaluates the clustering technique on an existing model for detecting prostate cancer [1][2]. The authors state the method provides more fine-grained information than saliency maps while capturing global behavior [2]. No quantitative benchmarks were included in the abstract or metadata released with the submission. The work was submitted to arXiv on November 18, 2025, and revised on May 27, 2026 [1]. The revision history indicates ongoing refinement, though the abstract remained unchanged between versions [1]. The paper falls under the Computer Vision and Pattern Recognition category on the preprint server [1]. Explainability in medical artificial intelligence remains an active research area. Regulatory bodies in multiple jurisdictions have signaled that model transparency will be a prerequisite for clinical deployment, though no binding standards have been finalized. The clustering method offers a departure from pixel-level attribution by summarizing model behavior in a format that pathologists can review as grouped tissue patterns rather than isolated heatmaps [2].