Abstract

562 Background: Newly diagnosed breast cancer specimens are routinely tested for oestrogen receptor (ER), progesterone receptor (PR) and human epidermal growth factor receptor-2 (HER2) status. This requires additional immunohistochemistry (IHC) +/- in situ hybridisation (ISH) assessment. This can increase laboratory/pathologist workloads and turnaround times. A computer-based approach could help to alleviate these issues. PANProfiler Breast is a deep-learning solution designed to predict ER/PR status and detect HER2 negative cases from whole slide images (WSIs) of haematoxylin and eosin (H&E)-stained breast cancer tissue. Here we present a blind validation of PANProfiler Breast. Methods: Three cohorts of H&E-stained archival breast cancers (400 cases, 692 WSIs) from St James’s University Hospital, Leeds, UK, were assigned for calibration (200 cases, 344 WSIs) and blind validation (200 cases, 348 WSIs). Slides were scanned at 40x magnification on an Aperio GT450 (Leica, Illinois, USA) scanner. PANProfiler returned “Positive”, “Negative” or “Indeterminate” for ER and PR, and “Negative” or “Indeterminate” for HER2. Results were compared to the ER/PR/HER2 status given in the corresponding pathology report. For calibration and blind validation, the WSI(s) corresponding to the IHC/ISH testing block was used for analysis. One case with missing HER2 status was excluded from HER2 analyses. Confusion matrices enabled analysis of concordance. While a conventional Positive/Negative confusion matrix was used for ER/PR, “Negative” results were used as the reference Positive event for HER2. Results: Case concordance for ER and PR status across the three cohorts ranged from 89.74-92.86% and 85.71-90.91% respectively. Additionally, sensitivity for ER ranged from 89.19-100%, with specificity ranging between 62.50-100%. For PR, sensitivity was between 87.50% and 93.48% and specificity between 75.00% and 83.33%. False negative rates for both ER and PR were between 0-10.81% and 6.52-12.50% respectively. Regarding HER2, sensitivity was relatively low, ranging between 30.77% and 37.04%. However, over 90% concordance was seen across all cohorts for cases PANProfiler predicted as HER2 “Negative”. Specificity and positive predictive value were also over 90% throughout the three cohorts. Additionally, false positive rates were 0.00%, 4.35% and 9.52% for HER2 across these same three cohorts. Test replacement rates varied amongst the cohorts with 70-84% for ER, 55-84% for PR and 22-27% for HER2. Conclusions: This blind validation of PANProfiler Breast shows impressive levels of performance for predicting ER and PR status in breast cancer WSIs. For HER2, the technology demonstrated a striking level of confidence for cases predicted as HER2 “Negative”. Further development and analysis are warranted to increase the number of HER2 negative BC WSIs that PANProfiler is able to identify.