Background: Eosinophil infiltration is a hallmark of eosinophilic esophagitis (EoE). Histologic assessment of esophageal biopsies is required for diagnosis and evaluation of treatment effect. Current diagnostic criteria are based mainly upon a peak eosinophil count (PEC) ≥15. Routine assessment of biopsies by pathologists to identify the area of greatest eosinophil density and maximal eosinophil count within a high power field (HPF) is the current gold standard. This process is laborious and subject to inter-observer variability. We aimed to develop a digital pathology analysis algorithm to assist pathologists with PEC assessment in digitized whole slide images (WSI) of esophageal biopsies in clinical trials.
Methods: Formalin-fixed paraffin embedded esophageal biopsies (N=45) from 45 patients with EoE representing a full spectrum of disease severity were sectioned and hematoxylin and eosin stained. Glass slides were scanned (40X) with a 510(k)-approved Aperio AT2 DX (Leica, Vista, California) imaging scanner at a CAP/CLIA accredited histology lab (AcelaBio, San Diego, California). WSI were manually annotated by a board-certified gastrointestinal pathologist using Visiopharm software version 2023.01 (Visiopharm®, Hoersholm, Denmark). Annotated images were a priori divided into training (n=20) and validation sets (n=25). A deep learning classification method trained an image analysis algorithm to detect and quantify eosinophils and create a heatmap to identify the epithelial area of greatest eosinophil density, thereby automating PEC assessment. Validation of algorithm performance was evaluated by (1) pathologist inspection of the output specificity (2) concordance between manual and automated PEC within the manually annotated HPF, where manual quantification was the gold standard and a concordance index (C-index) >0.75 was considered acceptable, and (3) precision of the algorithm to estimate PEC ≥15 per HPF.
Results: Manually annotated HPFs (n=20) in the training set included 2697 eosinophils. Algorithm outputs included eosinophil detection, an eosinophil density-based heatmap, and HPF with greatest eosinophil density (Fig 1). An accuracy of ≥90% for specificity of all algorithm outputs was confirmed by the pathologist. Excellent concordance (C-index=0.916, 95% confidence interval [0.849; 0.984]) for quantitation of PEC within HPFs using manual and automated methods was observed (Fig 2). The algorithm had 96% precision for histologic diagnosis of EoE based on a PEC threshold ≥15.
Conclusion: An algorithm was developed and validated for estimation of PEC within a HPF in digitized WSI of esophageal biopsies from patients with EoE. This algorithm may serve as a tool to augment assessment of histological disease activity in patients with EoE, increasing accuracy and reducing time associated with manual assessment.

<b>Figure 1</b>. Images of hematoxylin and eosin (H&E)-stained esophageal biopsy tissue sections are shown in panels A & B. Overlays of the original H&E-stained images generated by the digital pathology image analysis algorithm and used for automated analysis are shown in panel C, with the high-power field representative of the area of greatest eosinophil density shown in panel D.

Figure 1. Images of hematoxylin and eosin (H&E)-stained esophageal biopsy tissue sections are shown in panels A & B. Overlays of the original H&E-stained images generated by the digital pathology image analysis algorithm and used for automated analysis are shown in panel C, with the high-power field representative of the area of greatest eosinophil density shown in panel D.

<b>Figure 2</b>. Correlation between manual and automated peak eosinophil count within the same high powered field. PEC, peak eosinophil count.

Figure 2. Correlation between manual and automated peak eosinophil count within the same high powered field. PEC, peak eosinophil count.

Sean Gui

Appearances