A novel, computer-aided, scanning platform agnostic solution for grading carcinomas in breast biopsy whole slide images.
e12575 Background: The determination of breast cancer grade remains a problematic diagnostic issue in biopsies due to limited tissue sampling and the inherent inter-observer variability associated with the Nottingham (histologic) score. Significantly, biopsy grading corresponds only moderately with that based on excision specimens, with discordance rates of 21-30.0%, which is a concern for cases managed neoadjuvantly or with minimally ablative therapy. Although the advent of digital pathology has encouraged endeavors to automate breast cancer detection and grading, no method has yet received regulatory approval for clinical use or proven to be a platform agnostic solution. Methods: This study was the blinded clinical validation a novel, FDA breakthrough designation approved automated device-based predictive solution which uses a surrogate scale to determine breast carcinoma grade in core biopsies using H&E slide whole slide images (WSIs) only. Non-preselected malignant breast core biopsy clinical cases (n=173 WSIs; 107 cases) covering a broad spectrum of breast cancer morphological subtypes were scanned at x20 magnification on both on Aperio high-throughput T2/T3 systems (.SVS files) and Roche-Ventana DP200 scanners (.BIF files). The diagnostic gold standard reference was the reports of tertiary referral center breast subspecialty consultant histopathologists. Diagnostic outputs were also compared to case-matched resection specimen grades. Comparison of diagnostic outcomes with pathologists as gold standard was done using inter observer agreement (cohen kappa) with 95% CI and statistically significant chi-squared test p-value for two different scanner platforms. Results: Although biopsy-based diagnostic concordance with pathologists was very good, the device actually delivered a much better concordance between case-matched biopsy and resection specimen (accuracy: 95%, cohen kappa: 0.91) relative to pathologists’ assessments (accuracy: 80%). Conclusions: This indicates that the use of a single continuous surrogate grading scale is much less affected by limited tissue sampling than conventional ordinal morphological scales. This level of performance was independent of file format analyzed, demonstrating the device’s platform agnosia.