Abstract 167P
Background
As further de-escalation of axillary surgery is ongoing, new biomarkers that convey the same prognostic information as sentinel node status are called for.
Methods
To predict sentinel node status, we trained a deep learning (DL) image analysis model on H&E-stained whole slide images (WSIs) of primary tumors. For training, we used cases from the INSEMA standard arm (n=762, ca. 94 % HR+/HER2-, 3.5 % G3, less than 1% pT3/4, 13 % SLN-positive) and a cohort from the Women’s Clinic in Mannheim, Germany (n=150, all HR+/HER2-, G2, pT1/2, ca. 16 % SLN positive). We also fitted a logistic regression with clinical data (pT stage, Ki-67) for this task. Models were tested on a holdout INSEMA set (n=381), and the image model also on the higher risk TCGA BRCA cohort (n=650, ca. 72% HR+, ca. 55% SLN+). Vice versa, we trained a model on TCGA WSIs and tested it on the other cohorts.
Results
During training, the image and the clinical model yielded Areas under the Receiver Operating Characteristic Curve (AUROCs) of 0.62 and of 0.77 on the Mannheim WSIs, respectively. However, performance of the image model was random on the INSEMA (determined by blinded assessment) and TCGA BRCA test sets. The clinical classifier retained an AUROC of 0.62 on the INSEMA set. Inclusion of the image classifier output in the logistic regression did not improve performance on INSEMA. The image model trained on TCGA also yielded random performance on the INSEMA and Mannheim cohorts.
Conclusions
Image analysis algorithms trained on H&E stains of the primary tumors from INSEMA or TCGA using established techniques were unable to predict sentinel status, which may suggest a lack of systematic histological differences by lymph node status in these cohorts. Thus, DL-based WSI analysis may not be a good strategy to replace sentinel node assessment, especially in low- to intermediate-risk, hormone receptor-positive breast cancer.
Clinical trial identification
NCT02466737.
Legal entity responsible for the study
The authors.
Funding
TB, MS and EKH were funded by the TPI grant to TJB by the German Federal Ministry of Health. INSEMA trial is supported by German Cancer Aid (Deutsche Krebshilfe, Bonn, Germany), Grant No. 110580 and Grant No. 70110580 to University Medicine Rostock.
Disclosure
F. Marmé: Financial Interests, Personal, Invited Speaker: AstraZeneca, GSK/Tesaro, Clovis, Pfizer, Lilly; Financial Interests, Personal, Advisory Board: AstraZeneca, MSD, Novartis, Roche, Gilead/immunomedics, EISAI, PharmaMar, GenomicHealth, Myriad, Seagen; Financial Interests, Institutional, Invited Speaker: Seagen, Daiichi Sankyo, GSK, AstraZeneca, Roche, AstraZeneca, Novartis, Roche, Eisai, Gilead/Immunomedics, MSD, German Breast Group, AGO Research GmbH, Vaccibody, GSK; Financial Interests, Institutional, Advisory Board: Roche, Immunicom; Financial Interests, Institutional, Funding: AstraZeneca, Lilly, Seagen. C. Denkert: Financial Interests, Personal, Advisory Board: MSD Oncology, Daiichi Sankyo, Molecular Health, AstraZeneca, Roche, Lilly; Financial Interests, Personal, Invited Speaker: AstraZeneca, VmScope digital pathology software; Financial Interests, Institutional, Research Grant: Roche, Myriad, German Breast Group. V. Nekljudova: Financial Interests, Institutional, Full or part-time Employment: GBG; Financial Interests, Institutional, Research Grant: AbbVie, AstraZeneca, BMS, Daichi-Sankyo,Gilead, Novartis, Pfizer, Roche; Non-Financial Interests, Institutional, Writing Engagements: Daiichi Sankyo, Gilead, Novartzis, Pfizer, Roche, Seagen; Other, Institutional, Other, EP14153692.0: Patent; Other, Institutional, Other, EP21152186.9: Patent; Other, Institutional, Other, EP15702464.7: Patent; Other, Institutional, Other, EP19808852.8: Patent; Other, Institutional, Royalties: VM Scope GmbH. S. Loibl: Financial Interests, Institutional, Advisory Board, Member: Amgen, AstraZeneca, BMS, Celgene, EirGenix, GSK, Lilly, Pierre Fabre, Roche, Seagen, AbbVie, Sanofi, Gilead, Merck, Novartis, Relay Therapeutics; Financial Interests, Institutional, Invited Speaker: AstraZeneca, DSI, Novartis, Pfizer, Roche, Gilead, Seagen; Financial Interests, Institutional, Advisory Board: DSI, Pfizer, Olema; Financial Interests, Personal, Invited Speaker: Medscape; Financial Interests, Personal, Full or part-time Employment, CEO: GBG Forschungs GmbH; Financial Interests, Institutional, Invited Speaker, Ki67: VM Scope GmbH; Financial Interests, Institutional, Research Grant: AstraZeneca, Celgene, Novartis, Immunomedics/Gilead, Pfizer, Roche, Daiichi Sankyo; Financial Interests, Institutional, Funding: AbbVie, Molecular Health; Financial Interests, Personal, Other, PIPenelope/Padma: Pfizer; Financial Interests, Personal, Other, SC PALOMA3: Pfizer; Financial Interests, Personal, Other, SC SOLAR1: Novartis; Financial Interests, Personal, Other, SC ASCENT: Immunomedics/Gilead; Financial Interests, Personal, Other, SC HERCLIMB: Seagen; Financial Interests, Personal, Other, SC Katherine: Roche; Financial Interests, Personal, Other, SC Capitello; EC Cambria 1: AstraZeneca; Financial Interests, Personal, Other, SC Inavo: Roche; Financial Interests, Personal, Other, SC Destiny B05; SC Destiny B09: Daiichi Sankyo; Non-Financial Interests, Principal Investigator, After publication of primary endpoint: PI Aphinity; Non-Financial Interests, Advisory Role, Group in Germany responsible for breast cancer guidelines: AGO Kommission Mamma; Non-Financial Interests, Member, German Gynaecological Oncology society: AGO; Non-Financial Interests, Member, German Cancer Society: DKG; Non-Financial Interests, Member: ASCO; Non-Financial Interests, Member, Member guideline committee; past chair in ESMO Breast: ESMO; Other, EP14153692.0No financial interest, Institutional: Patent; Other, EP21152186.9No financial interest, institutional: Patent; Other, EP15702464.7No financial interest, institutional: Patent; Other, EP19808852.8 No financial interest, Institutional: Patent. All other authors have declared no conflicts of interest.