Poster S1 - Deep Learning Enables Virtual Histological Staining of Label-free Autopsy Tissue Sections

Introduction: Autopsy examination is crucial for determining the causes and manner of death. Initially, a naked-eye-based gross examination is conducted, followed by the microscopic histological analysis, which is pivotal in revealing cellular and molecular changes associated with death progression. The histological analysis not only confirms the causes of death but also provides insights into disease development and the efficacy of any medical treatments. However, traditional histochemical staining of autopsy samples frequently encounters the problem of suboptimal staining quality. Unlike biopsies with immediate tissue fixation, autopsies usually experience a delay from the time of death to the operation of the autopsy and tissue fixation. This delay leads to autolysis, a self-digestion process that occurs in the deceased's body, which negatively influences the binding of chemical dyes to biomolecules within the tissue, consequently leading to staining artifacts. Additionally, the conventional histochemical staining process is labor-intensive, costly, and time-consuming. To address these challenges, we introduce a deep learning-based virtual staining technique (Fig. 1a) for label-free autopsy tissue samples, which can effectively mitigate autolysis-induced artifacts while saving a large amount of time, labor, and reagents.

Materials and

Methods: Our presented autopsy virtual staining technique leverages two autofluorescence microscopic images (DAPI and TxRed channels) from unlabeled autopsy tissue sections as inputs. These images are digitally transformed into their bright-field hematoxylin and eosin (H&E) stained equivalents using a trained neural network, which was built using a custom-designed data-efficient framework named RegiStain. This training framework integrates an image registration network into the training process of the virtual staining network, progressively achieving the precise pixel-level alignment between the network’s outputs and their histological ground truth images. After being trained and validated with >16,000 microscopic image pairs, the virtual staining network model was blindly tested on ~2000 image fields of view, each with 8000 × 8000 pixels, obtained from 10 unique patients.

Results, Conclusions, and Discussions: Compared to histochemically stained images from severe-autolysis regions, where traditional staining exhibits numerous missing nuclei, our virtual staining method significantly enhances the contrast of nuclear details and cellular morphology as shown in Fig. 1b. To validate the effectiveness of our method, we also compared virtually stained H&E images with their histochemically stained counterparts from well-preserved regions, where the actual high-quality ground truths are accessible. In these well-preserved regions, a high concordance was observed between virtual and histochemical staining results in various aspects such as nuclear, cytoplasmic, and extracellular details. The subsequent quantitative analysis using digital image processing algorithms and evaluations by board-certified pathologists both showed there is no statistically significant difference in staining quality between the virtual and traditional staining methods in these well-preserved samples. Moreover, our deep learning-based virtual staining model can generate virtual H&E images in < 2 seconds per mm² on a single-GPU computer, offering substantial time savings compared to traditional H&E staining, which typically requires 2-3 hours to process a tissue sample. In summary, we introduced a novel post-mortem virtual staining technique, serving as a rapid, cost-effective, and resource-saving histological staining solution that consistently yields high-quality results, even in severe-autolysis regions where the traditional histochemical staining falls short. It will be especially valuable in urgent situations, such as during pandemics, where quick and efficient autopsy analysis is necessary. Furthermore, this technique has the potential to be applied beyond autopsies, including staining of necrotic tissues in biopsies, which encounter similar challenges to autolytic tissues in histochemical staining.

Acknowledgements (Optional):