308P - Regression-based deep-learning predicts breast cancer recurrence risk score from pathology slides

Background

Breast cancer (BC) treatment decisions can be challenging, particularly in patients with early-stage hormone receptor-positive/HER2-negative (HR+/HER2-) tumors. Prognostic assays, such as the 21-gene recurrence score assay (Oncotype DX), are used to determine individual patient risk of recurrence and guide decision on whether to administer adjuvant chemotherapy or not. However, these tests are costly and time-intensive to perform. Deep-Learning (DL) can predict molecular biomarkers from routine hematoxylin and eosin (H&E) pathology slides, potentially serving as an inexpensive and accessible pre-screening tool. We aim to use DL for the outcome prediction of the Oncotype DX test.

Methods

We trained a DL-based regression model in a weakly-supervised manner to detect Oncotype DX risk score directly from routine H&E-stained pathology slides in a large (n=5,303) cohort of early-stage HR+/HER2- BCs from Memorial Sloan Kettering (MSK). We then externally validated the model onto The Cancer Genome Atlas (TCGA) BC cohort for which Oncotype DX scores were available (n=100). The model was evaluated using the Pearson’s correlation coefficient r. Moreover, the recurrence score was divided into low (<11) and intermediate/high (≥11) risk groups, enabling evaluation via the area under the receiver operating characteristic curve (AUROC).

Results

The model yielded a significant r of 0.59 (p<0.0001) on the test set of the training cohort of MSK BCs and 0.58 (p<0.0001) on the external cohort of TCGA BCs. When binarizing the predictions into intermediate/high and low recurrence risk groups, we found that the model was particularly accurate for identifying patients with intermediate/high risk scores, reaching an AUROC of 0.87 (± 0.04) and 0.79 (± 0.03) in the MSK and TCGA cohort, respectively.

Conclusions

We developed a DL-model that can accurately predict Oncotype DX score from routine pathology slides, offering a cost-effective and time-efficient pre-screening tool for identifying a high-risk subgroup HR+/HER2- early BCs and to democratize the access to complex but clinically useful biomarkers. This approach could support informed treatment decisions by accurately determining recurrence risk, ultimately improving patient outcomes.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

TU Dresden.

Funding

EKFZ.

Disclosure

J.S. Reis-Filho: Financial Interests, Personal, Other, Consultant: Goldman Sachs, Eli Lilly, Saga Diagnostics; Financial Interests, Personal, Other, Member of the Scientific Advisory Board and Consultant: Repare Therapeutics, Paige.AI; Financial Interests, Personal, Advisory Board: Personalis, Roche Tissue Diagnostics; Financial Interests, Personal, Advisory Board, Member of the Scientific Advisory Board: Bain Capital; Financial Interests, Personal, Advisory Board, Ad hoc member of the Pathology Scientific Advisory Board: Daiichi Sankyo, Merck; Financial Interests, Personal, Advisory Board, Ad hoc member of the Oncology Scientific Advisory Board: AstraZeneca; Financial Interests, Personal, Advisory Board, Member of the SAB: MultiplexDX; Financial Interests, Personal, Member of Board of Directors: Odyssey Bio, Grupo Oncoclinicas; Financial Interests, Personal, Stocks/Shares: Repare Therapeutics; Financial Interests, Personal, Other, Stock options: Paige.AI. J.N. Kather: Financial Interests, Personal, Invited Speaker: Fresenius, Eisai, MSD; Financial Interests, Personal, Advisory Board: Owkin, DoMore Diagnostics, Panakeia, London, UK. All other authors have declared no conflicts of interest.