Abstract:
Formalin-fixation and paraffin-embedding (FFPE) is a specimen preparation and preser vation technique that has been used in histology and pathology since the late 19th cen tury. Because the preparation of FFPE specimens is a complex, lengthy, and difficult to standardize process, and due to the complex histological and cytological characteristics of tissue, FFPE slides often contain defects. Defects arise during tissue fixation, pro cessing, embedding, microtomy, staining, and coverslipping. These defects are referred to in images as artifacts, a term which encompasses staining inconsistencies, tissue folds, chattering, pen marks, blurring, air bubbles, and contamination, in addition to some other defects. We propose an unpaired image-to-image translation approach, FFPE++, which corrects artifacts in FFPE slides for digital pathology. Our method is a deep-learning-based approach which uses contrastive learning with spatial atten tion block and self-regularization loss, leading to higher quality in terms of both the visibility of textural details and cellular features. 10 board certified pathologists have performed comparative tests between our FFPE++ method and the standard FFPE sections of the ovary, thyroid, and lung, showing that our approach results in a visually coherent images for histopathological diagnosis.|Keywords : : formalin-fixed paraffin-embedded, deep learning, unpaired image-to-image translation, generative adversarial networks, digital pathology, histological artifacts, histopathological image analysis.
