Reducing technical variability in image analyses of haematoxylin and eosin stained tissues

<ns3:p>&lt;br&gt;&lt;b&gt;Objective:&lt;/b&gt; The aim of this study is to develop and evaluate a preprocessing pipeline aimed at reducing technical variability in histomic features extracted from whole-slide images (WSIs) of endometrial cancer tissue using foundation models, thus improving the reliability and generalizability of downstream computational pathology analyses.&lt;/br&gt;&lt;br&gt;&lt;b&gt;Methods:&lt;/b&gt; Haematoxylin and eosin (H&amp;E) stained images from three datasets (TCGA UCEC, CPTAC UCEC and Cracow UCEC) were preprocessed using a pipeline that included filtration of artifacts and Vahadane-stain normalisation. Four histopathological foundation models (ResNet18 Histo, CTransPath, UNI, H-optimus-0) were used as feature extractors. Batch effects were evaluated before and after preprocessing using UMAP visualization plots and LISI and ARI metrics.&lt;/br&gt;&lt;br&gt;&lt;b&gt;Results:&lt;/b&gt; For each foundation model tested a strong batch effect was observed. The proposed method allowed for the reduction of technical batch effect arising from differences between datasets (e.g., for ResNet18 Histo, ARI decreased from 0.7780 to 0.1428, and LISI increased from 0.0043 to 0.1393), but it did not reduce variability related to the tissue-embedding medium for most models. Additionally, biological diversity associated with tumour grade decreased slightly (e.g., for UNI, ARI decreased from 0.0391 to 0.0384, and LISI increased from 0.5658 to 0.6485).&lt;/br&gt;&lt;br&gt;&lt;b&gt;Conclusions:&lt;/b&gt; The introduced preprocessing approach effectively mitigates technical variability in histomic features extracted by foundation models, improving the comparability of WSIs from different sources. However, differences due to the embedding medium remain challenging. The findings emphasize the need to address batch effects in computational pathology and suggest that robust preprocessing is a crucial element of data preparation for model building.&lt;/br&gt;</ns3:p>