Deep learning chronic wasting disease (CWD) immunohistochemistry (IHC) image dataset

The dataset contains 143 whole-slide images (WSI) containing a combination of central nervous system tissue, typically obex containing the dorsal motor nucleus of the Vagus (DMNV; n= 137) and retropharyngeal lymph nodes (RPLN; n = 114) derived from surveillance diagnostic samples and farmed cervid depopulations. Species represented in the training data set included white tailed deer (n = 68), sheep (n= 54), elk (n = 14), goat (n = 4), and moose (n = 3). Of the 143 slides, 54 were identified as suspect (i.e. detected) and 89 were not detected. Ground truth annotations for lymphoid follicles in retropharyngeal lymph nodes and the dorsal motor nucleus of the Vagus (DMNV) in obex samples were manually annotated by a transmissible spongiform encephalopathy (TSE) trained board-certified veterinary anatomic pathologist. Annotations were performed in QuPath 5.0 using the brush tool. In total, the training data set contains 3,296 annotations broken down into +/- DMNV regions (n = 224+/438-, respectively) and +/- lymphoid follicular regions (n = 1295+/1339-, respectively). The dataset was collected and annotated in order to train deep neural networks for tissue type and anatomical structure detection. The code is available in an accompanying Github repository.

Methods
Dataset case selection

Formalin-fixed, paraffin-embedded (FFPE) tissues, including retropharyngeal lymph node and obex submitted for TSE surveillance were retrospectively selected from the Washington Animal Disease Diagnostic Laboratory (WADDL) as well as United States Department of Agriculture Agricultural Research Service Animal Disease Research Unit (USDA-ARS-ADRU) scrapie research cases. Inclusion criteria required that cases had been previously evaluated by a TSE trained veterinary pathologist and assigned one of the following diagnostic categories based on immunohistochemistry (IHC): Detected, Not Detected, Location, or Insufficient Follicles. These categories reflect standard interpretive outcomes used in TSE surveillance programs and represent the full spectrum of tissue and staining conditions encountered in diagnostic practice. Cases were excluded if they had been assigned an unacceptable or unsuitable diagnostic code typically due to poor sample fixation and subsequent postmortem autolysis.

Tissue processing and staining

All immunohistochemical staining and evaluation were performed according to the NVSL document: Detection of Scrapie and Chronic Wasting Disease by Immunohistochemistry.22 Briefly, all samples were formalin fixed in 10% neutral buffered formalin with routine processing and embedding. Slides were cut at 4 µm thickness and mounted on NVSL approved charged slides. Mounted slides were pretreated with 96% Formic Acid for 5 minutes and transferred to Tris Buffer 0.05M, pH 7.5 to rinse. Decloaking was performed utilizing the Diva Decloaker solution (Cat# DV2004; Biocare Medical, Pacheco, CA) in combination with the BioCare antigen retrieval chamber pot (Cat# DCARC0001; Biocare Medical, Pacheco, CA).

IHC was performed on the Ventana Discovery Ultra autostainer (Roche Diagnostics, Indianapolis, IN), employing the ready-to-use (RTU) dispensers for F99 monoclonal antibody provided in the Anti-Prion Research Kit (Cat# 760-231; Roche Diagnostics, Indianapolis, IN). Individual RTUs additionally contain biotinylated secondary antibody, an alkaline phosphatase–streptavidin detection system, and a substrate chromogen composed of fast red A, naphthol, and fast red B, followed by hematoxylin counterstaining. Each staining run included a positive control section.

Digitization

All whole slide images were acquired using a Leica GT450 using default settings at a magnification of 400x (image resolution: 0.264 µm/pixel or 96,154 pixels/inch). All WSI images were saved as .SVS files and viewed or annotated using QuPath throughout the project.