Validating an AI-based analytic tool for IHC staining QA: precision studies of the digital pathology pipeline
Recommended Citation
Tuthill J. Validating an AI-based analytic tool for IHC staining QA: precision studies of the digital pathology pipeline. Virchows Arch 2024; 485:S398.
Document Type
Conference Proceeding
Publication Date
9-6-2024
Publication Title
Virchows Arch
Abstract
Background & objectives: Standardization of immunohistochemistry staining quality assessment is critical for diagnostic accuracy. Pathologists currently assess stain quality subjectively, comparing control sections to patient tissue. Qualitopix (Visiopharm, Denmark), a cloud-based artificial intelligence platform for uses quantitative analysis for scoring stained slides. Methods: Glass slides were produced from two 4-core standardized cell-line blocks (Histocyte Laboratories, Newcastle, England) with epitopes for estrogen receptor (ER) and progesterone receptor (PR) of increasing intensities, stained using Ventana Benchmark Ultra and scanned on DP 200 and HT scanners (Roche, Basel, Switzerland). An intra-scanner precision study was performed by comparing Qualitopix-derived H scores of ER and PR slides. Results: Intra-scanner precision studies demonstrated consistent reproducibility using both scanners: %CV for ER cores were 0%, 10.7%, 2.3% and 0.3% for cores 1 [0 +/- 0.003], 2 [0.3 +/- 0.25], 3 [2.4 +/-1.4] and 4 [79 +/- 0.65] respectively. % CV for PR cores were 0%, 0.6%, 0.4% and 0.1% for cores 1 [0 +/- 0.01], 2 [29 +/-4.5], 3 [65+/-2], and 4 [94+/-2] respectively. Concordance studies revealed tight agreement. ICC for ER cores were 0.64 (moderate), 0.95(excellent), 0.95(excellent) and 0.68 (moderate) for cores 1, 2, 3 and 4 respectively. ICC for PR cores were 0.64 (moderate), 0.87(good), 0.96(excellent) and 0.68 (moderate) for cores 1, 2, 3 and 4 respectively. Conclusion: Quality assurance is essential to the use of digital pathology, particularly to the application of AI. Studies of precision, reproducibility, and accuracy are lacking in the literature. This study demonstrates the precision characteristics of one vendor's digital pathology product line.
Volume
485
First Page
S398
