Deep Learning and Digital Pathology Power Precision Treatment of Triple-Negative Breast Cancer

Abstract

Background: Molecular stratification and tailored target therapy bring clinical benefit for patients with triple-negative breast cancer (TNBC), but it is difficult to implement comprehensive molecular testing in clinical practice. Thus, we aim to devise an approach based on digital pathology and deep learning for the molecular and prognostic stratification of TNBC. Methods: We collected digital whole slide images (WSIs) (N=425) of our previously established TNBC cohort with multi-omics data. A deep learning-based workflow was developed and applied to these WSIs to train and validate neural network models to predict multi-omics molecular features, to identify molecular subtypes, and to improve prognostic evaluation. The models that showed high prediction accuracy were further validated on the TNBC cases from TCGA (N=143). Findings: A model was first developed for automatic tissue type classification, which enabled selection of certain tissue types on WSIs for the following prediction. Numerous molecular features can be inferred from WSIs including the somatic PIK3CA mutation, germline BRCA2 mutation and PD-L1 expression. All of the four molecular subtypes of TNBC can be accurately identified based on WSIs and distinctive morphological patterns were revealed for each subtype. The addition of image features to clinical prognostic factors significantly improved the accuracy of relapse risk assessment. The models for predicting PIK3CA mutation, PD-L1 expression, TNBC subtypes and relapse risk can be well generalized to the TCGA TNBC cases. The complete prediction workflow along with validated models was modularized and deployed on an online platform, which can realize real-time one-stop prediction for newly-uploaded WSIs. Interpretation: We proposed a deep learning-based workflow and developed neural network models to predict clinically relevant information of TNBC from pathological WSIs. Our findings and established platform may enable the implementation of artificial intelligence guided precision treatment in clinical trials and future routine practice. Funding: Fudan University Shanghai Cancer Center. Declaration of Interest: None to declare Ethical Approval: The tissue samples used in our study were obtained after the approval from the FUSCC Ethics Committee.