e12517 Background: HER2 amplification or overexpression is negative prognostic factor that occurs in circa 15% of primary invasive breast cancers, with positive results opening up eligibility for HER2 targeted (e.g. trastuzumab) combination therapy. However, HER2 intratumoral heterogeneity has been proposed as an explanation for the development of resistance to anti-HER2 targeted therapy. The aim of this retrospective study was to test the prognostic value of a HER2 heterogeneity index on ER+ breast cancer H&E whole slide images (WSIs). Methods: ER+ breast cancers slides ( n=60 cases) of known HER2 status (immunohistochemically assessed over the 0 to 3+ range, with HER2 2+ borderline cases validated by fluorescence in situ hybridization) were digitized on an Aperio AT3 scanner. A subset of these patients ( n=29) was accounted for by HER2 amplified/borderline cases. A HER2 heterogeneity index was tested and compared to standard clinical HER2 status in these cases and related to progression free survival (PFS). Survival analyses were performed using Kaplan-Meier (KM; with median cut-off) and Cox Proportional Hazards (as a continuous variable) models. Results: The HER2 heterogeneity index performed better in terms of prognostication than clinical HER2 status alone in relation to PFS (KM P=0.019; HR 8.06, P=0.01 vs. KM P=0.02; HR 6.75, P=0.048). These differences were more pronounced in the HER2 amplified/borderline tumor subset, where the HER2 heterogeneity index outperformed clinical HER2 status (KM P=0.04; HR 6.38, P=0.001 vs. KM P=0.10; HR 0.64, P=0.20). Conclusions: These findings highlight the merit of identifying HER2 heterogeneity from H&E slides alone and establish this index as a robust prognostic factor in determining PFS in ER+ breast tumors. Future work will extend this investigation into ER- tumors and evaluate patient-specific tumor responses to anti-HER2 targeted therapy.

e12518 Background: Despite its purported prognostic significance in breast cancer, Ki67 index assessment remains poorly standardized and features high discordance rates amongst pathologists. Moreover, its clinical utility is presently confined to low stage, ER+/HER2- tumors in determining the advisability of adjuvant chemotherapy. Unfortunately, established cut-offs limit its utility to tumors with either high or low Ki67 index extremes. While Ki67 is a marker of actively dividing cells, it fails to capture the detail of cell cycle phase and dynamics which could be informative in terms of prognosis and tumor sensitivity to specific therapies (e.g. CDK4/6 inhibitors). This study therefore developed two novel indices as surrogates of (i) Ki67 index and (ii) quiescent cell population load (QPL). Methods: Breast cancer (comprising evenly distributed hormone receptor/HER2 status cases) H&E slides ( n = 79 cases/108 slides) were digitized on an Aperio DT3 scanner, and used to extract surrogate Ki67 and QPL indices. Sections were also stained for Ki67 by immunohistochemistry (IHC) using a clinically validated assay and digitized. Whole slide image (WSI) tumor Ki67 counts were performed on QuPath and used to validate the surrogate Ki67 index (Cohen’s kappa score). Indices (i) and (ii) were related to progression free survival (PFS). Survival analyses were performed using Kaplan-Meier (KM; with median cut-off) and Cox Proportional Hazards (as a continuous variable) models. Results: The surrogate Ki67 index showed good concordance with IHC scores (kappa = 0.76; 95%CI 0.61-0.91; P< 0.00001). However, this surrogate index performed better as a prognostic indicator of PFS compared to conventional Ki67 IHC (KM P = 0.048; HR = 1.35, P = 0.015 vs. KM P = 0.70; HR = 1.23, P = 0.08). Prognostically, the QPL index outperformed both Ki67 indices (KM P = 0.03; HR = 3.22, P = 0.001). Conclusions: We have developed two novel surrogate indices of Ki67 and QPL that can be readily automated to analyze H&E breast cancer WSIs. Our results show that both outperformed conventional Ki67 IHC evaluation in terms of prognostication, applied across molecular subtypes, improved informativeness of mid-range Ki67 index calls, and could potentially have predictive merit in selecting patients for cell cycle targeted therapies such as CDK4/6 inhibitors.

593 Background: Digital pathology has fostered the development of automated diagnostic solutions. However, current technologies in breast cancer remain unable to determine HER2 amplification status, which is established by immunohistochemistry (IHC) and/or fluorescent in situ hybridization (FISH). These ancillary tests carry a significant cost, prolong diagnostic time and fail to capture HER2 tumor heterogeneity and tumor infiltrating lymphocyte (TIL) burden, both of which determine the effectiveness anti-HER2 targeted therapy. Methods: This study describes the real-life clinical context development and validation of a patented novel, universal, automated, white-box, scanning platform agnostic solution that determines HER2 amplification status, prognostically significant TIL levels and tumor heterogeneity index (HI) from hematoxylin and eosin (H&E) stained malignant breast biopsy whole slide images (WSIs) alone. Unlike conventional artificial intelligence-based approaches, the underlying proprietary algorithm’s prediction criteria are explainable and are based on deterministic, hard-coded observational relationships of scale constructed from image morphological features mapped to observables representing underlying tumor-related perturbations in biological pathways/mitotic checkpoints.This includes G1/S deregulation signatures reflecting oncogenic HER2-neu. Results: Blinded validation of HER2 status prediction (n = 197 WSIs; 118 independent cases/patients) showed excellent diagnostic performance (κ = 0.85) relative to existing standard-of-care methodologies. This was independent of WSI file format, background histology, tumor subtype/grade or hormone receptor status. The device also displayed good accuracy (92%) in determining TIL profiles, and its combined HER2-TIL and HER2-HI prediction scales both exhibited a significant association with progression-free survival (CoxPH hazard ratio: 1.856; 95% CI: 1.002-3.438; p = 0.049 and CoxPH hazard ratio: 3.45; 95% CI: 0.95-12.55; p = 0.060). Conclusions: This technology opens up the future possibility of bypassing existing ancillary HER2 profiling investigations, thus potentially reducing laboratory workloads/healthcare costs while accelerating diagnostic turnaround times for patient benefit. In the interim, if used as an adjunct tool, this device could provide an objective HER2 testing reference scale while the robustness of its prediction of patient response to anti-HER2 targeted therapy is fully explored.

Background Breast cancer patients with estrogen receptor (ER)+/HER2- (and usually node-negative) tumors can avail themselves of Oncotype DX Breast Recurrence Score (ODXRS) testing to predict their risk of distant recurrence within 9 years and, consequently, putative chemotherapy benefit. However, ODXRS testing requires sufficient tumour availability and specimen shipping, which imposes time and financial burdens to testing which have to be met by healthcare systems. The advent of digital pathology offers a potential avenue for exploring computer-aided diagnostic solutions which may overcome these hurdles by extracting the requisite information from hematoxylin and eosin (H&E)-stained tissue whole slide images (WSIs) alone. In turn, this technology could significantly reduce diagnostic turnaround times and cost, and improve accessibility and test reproducibility, thereby enabling healthcare systems to run more efficiently and offer patients more timely results. Ideally, such a platform should incorporate a measure of the underlying tumor biology to provide a fully explainable, white box solution, and may offer further insights into the identification of early recurrence events. Aims The aim of this study was to establish whether our computer-aided solution’s (Q-Plasia OncoReader Breast, QPORB) digital biomarker representing G1/S cell cycle deformations extracted from H&E WSIs was prognostic for disease-free survival (DFS) and could predict disease recurrence, particularly in the setting of low risk ODXRS breast cancers. Methods Primary breast cancer resection/excision specimens (n=70 cases) sent for ODXRS testing from St James’s University Hospital, UK (2016-2019) were collected. Anonymised diagnostic glass slides (n=198 slides) of H&E-stained tumors were scanned at x20 magnification on an Aperio AT2 scanner. In parallel, relevant clinical and histological data were collected from pathology reports and electronic patient records, including both ODXRS and recurrence events during follow-up. The QPORB recurrence scale (QPORB-RS), which combines statistical physics and tumor biology to identify image-based malignant cell cycle deformation, extracts prognostic information from WSIs. The contribution of potential confounders (age, stage, grade, lesion size, Nottingham prognostic index and Charlson score) were accounted for. Results The QPORB-RS was prognostic for DFS for patients with predominantly node-negative (including node micro-metastases) HR+/HER2- tumors over a median follow-up period of 5 years (P=0.02; dichotomized Kaplan Meyer with median cut-off). The QPORB-RS concurred with ODXRS’s high vs. low recurrence risk in 73% (19/26) and 61% (27/44) of cases, respectively, with an overall agreement of 66% (46/70). Moreover, the QPORB-RS identified all 5 patients who had recurrences (with ODXRS of 6, 9, 10, 21 and 26, and ages of 55, 66, 42, 35 and 50 years, respectively) as being high risk in the subset of those given a low (including historically intermediate) ODXRS and who did not receive chemotherapy. Conclusion The QPORB-RS is a good prognostic test of risk of disease recurrence in breast cancer patients with predominantly node-negative (including node micro-metastases) HR+/HER2- tumors within a median 5-year follow-up period. Our efforts are now focussed on extending this cohort and establishing the prognostic value of the QPORB-RS across all breast carcinomas, regardless of molecular subtype, stage/node positivity and menopausal status. Citation Format: Satabhisa Mukhopadhyay, Tathagata Dasgupta, Elizabeth Walsh, Rebecca Millican-Slater, Andrew hanby, Joanne Stephenson, Craig A. Bunnell, Nicolas M. Orsi. Prediction of disease recurrence in low risk Oncotype Dx breast cancers from digital H&E-stained whole slide images of pre-treatment resections alone [abstract]. In: Proceedings of the 2022 San Antonio Breast Cancer Symposium; 2022 Dec 6-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2023;83(5 Suppl):Abstract nr P3-05-48.

e12535 Background: Categorical combinations of ER, PR, HER2, and Ki67 levels are traditionally used to classify patients into luminal A and B-like subtypes in order to inform treatment choice. Accounting for nearly 70% of all breast malignancies, luminal cancer is heterogeneous, harboring subtypes with distinct molecular profiles and clinical outcomes. Although most patients with luminal-type disease respond well to endocrine therapy alone, some develop recurrences benefiting from additional cytotoxic therapy. Identifying such cases a priori remains a challenge but would enable patients to be spared the debilitating side-effects of ineffective chemotherapy. In this regard, the efficacy of chemotherapy and disease recurrence relate to (i) ER driven G1/S perturbations and/or (ii) quiescent cell populations arrested in the G0/G1 phase of cell cycle. This study aimed to develop a histopathology whole slide image (WSI)-based, low cost, rapid and automated approach to: (i) predict ER/PR/Ki67 status, (ii) quantify quiescence burden, (iii) develop a G1/S-based patient stratification system for luminal A/B patients, and (iv) achieve a quiescence burden-based stratification of TNBC patients. Methods: This investigation centered on the initial clinical validation of a novel, immunostaining-free technology which uses information extracted from pre-treatment hematoxylin and eosin (H&E) stained slide WSIs alone to achieve these aims. Unlike conventional artificial intelligence-based approaches, the underlying proprietary algorithm and its prediction criteria are based on deterministic, hard-coded observational relationships of continuous scales drawn from WSI morphological features. In this instance, these represent tumor-related biological pathway disruptions and mitotic checkpoint perturbations, where G1/S perturbations enable luminal subtype stratification, and G0/G1 perturbations reflect quiescence burden. Back projecting the algorithm’s quiescence burden output on to the original WSIs enables morphological patterns to be mapped to quiescence burden.