Efficient Special Stain Classification

TL;DR

This study tackles automating the classification of 16 stain classes, including 14 common special stains and H&E variants, to improve metadata quality in digital pathology. It compares a computationally intensive patch-based Multi-Instance Learning approach with ABMIL against a lightweight thumbnail-based method, evaluating both internal (TUM) and external (TCGA) data. Internal results favor MIL for maximum slide-level precision, while external generalization and throughput strongly favor the thumbnail classifier, which achieves orders-of-magnitude higher throughput and better domain robustness. The findings suggest a practical trade-off: patch-based methods excel in fine-grained discrimination, but thumbnail-based classification provides a scalable, robust, and deployment-friendly solution for routine quality control in institutional digital pathology workflows, with potential extensions to uncertainty estimation and open-set handling.

Abstract

Stains are essential in histopathology to visualize specific tissue characteristics, with Haematoxylin and Eosin (H&E) serving as the clinical standard. However, pathologists frequently utilize a variety of special stains for the diagnosis of specific morphologies. Maintaining accurate metadata for these slides is critical for quality control in clinical archives and for the integrity of computational pathology datasets. In this work, we compare two approaches for automated classification of stains using whole slide images, covering the 14 most commonly used special stains in our institute alongside standard and frozen-section H&E. We evaluate a Multi-Instance Learning (MIL) pipeline and a proposed lightweight thumbnail-based approach. On internal test data, MIL achieved the highest performance (macro F1: 0.941 for 16 classes; 0.969 for 14 merged classes), while the thumbnail approach remained competitive (0.897 and 0.953, respectively). On external TCGA data, the thumbnail model generalized best (weighted F1: 0.843 vs. 0.807 for MIL). The thumbnail approach also increased throughput by two orders of magnitude (5.635 vs. 0.018 slides/s for MIL with all patches). We conclude that thumbnail-based classification provides a scalable and robust solution for routine visual quality control in digital pathology workflows.

Figures (7)

• Figure 1: Overview of the two classification pipelines. (a) The slide is segmented, divided into patches, and patch features are extracted and aggregated by ABMIL for prediction. (b) The whole slide thumbnail ($896 \times 1792$ px) is processed in a single forward pass, avoiding tissue segmentation and patch-level feature extraction.
• Figure 2: Confusion matrices for thumbnail and MIL ($k=\text{all}$) approaches on the TUM test set (fine, 16 classes). Values below 0.01 are omitted for clarity.
• Figure 3: Representative $40\times$-magnification tiles of Von Kossa and Prussian Blue showing the sometimes used pink counterstain in Von Kossa, identical to the counterstain used in the Prussian Blue staining protocol.
• Figure 4: Interpretability maps for an Alcian Blue–PAS stained slide. (a) Thumbnail GradCAM. (b) MIL attention weights map. (c) Patch-level predictions map.
• Figure 5: Ablation studies. (a) Macro F1 score vs. patch magnification for patch-level and MIL ($k{=}\text{all}$) classifiers. (b) Macro F1 score vs. thumbnail input resolution. Shaded areas show the 95% CI across 5-fold cross-validation.