Staining condition visualization in digital histopathological whole-slide images

Staining condition is one of the essential properties in digital pathology for developing computer-aided diagnosis (CAD) systems; however, it is challenging to analyze the staining condition of giga-pixel whole-slide images (WSIs) due to the high data volume. In this study, we proposed an intuitive method to visualize the color style of Hematoxylin and Eosin (H&E) stained WSIs, which is scalable to large real-world cohorts. For this, representative color spectrums are obtained by K-means clustering on slide-level, and the pair-wise distance between spectrums is formulated as a matching problem. Lastly, we use multi-dimensional scaling (MDS) algorithm to obtain 2-dimensional embeddings for WSIs, which are suitable for visualization. We validated the method on lung adenocarcinoma cases and lung squamous-cell carcinoma cases in The Cancer Genome Atlas (TCGA) program. Through the well-visualized staining pattern map, slides with low staining quality or with abnormal staining conditions can be easily recognized. Furthermore, we give a demo usage of the proposed method in the context of a lung cancer segmentation task. Our main conclusions including, (1) biases in staining pattern distribution will harm the performance of CAD systems; (2) weakly stained slides are more challenging than heavily stained slides; (3) stain augmentation can deal with a certain level of staining variation, but not all of it; (4) light stain augmentation can generate more realistic training samples.

Automated summary

This study proposes an intuitive method to visualize the color style of Hematoxylin and Eosin (H&E) stained whole-slide images (WSIs) and validates its effectiveness in lung cancer research.