Deep learning predicts the effect of neoadjuvant chemotherapy for patients with triple negative breast cancer

Background

Triple negative breast cancer (TNBC) is an aggressive subcategory of breast cancer with poor prognosis and high risk of recurrence after treatment. In a subset of cases systemic chemotherapy is offered before surgery, so called neoadjuvant chemotherapy (NAC), to downstage the disease resulting in 40–50% of cases to a pathological complete response. Meanwhile, patients receiving NAC suffer from toxic side effects and in a proportion of patients a significant amount of residual tumor remains. This study aims to predict the outcome of NAC with deep learning technology based on the microscopic morphological characteristics in whole slide images of hematoxylin and eosin (H&E) slides from the pre-operative tumor biopsy before chemotherapy.

Methods

A convolutional neural network was trained on 221 H&E-stained biopsies of carcinoma of no special type from 205 patients scanned at 40×. Cases were divided in three cohorts, with a good, moderate, or bad response to NAC based on the EUSOMA scoring according to the pathology report of the subsequent tumor surgery specimen. We defined good, moderate, and bad response as residual tumor <10%, 10–50%, and >50%, respectively. Manual segmentation of the tumor area was performed comprising invasive carcinoma with a small rim of surrounding benign tissue. The model was tested on 52 new biopsies of 50 patients. Because of the relative low number of moderate and bad responder cases, and to achieve a better discrimination for potential visual biomarkers, the moderate and bad response cohorts were merged.

Results

The predictive performance of the model was calculated by means of the area under the receiver operator curve (AUC ROC). 95% Confidence intervals (CIs) were calculated for better understanding of the range of values. In the test set, the AUC ROC performance score was 0.696 with a CI of 0.532–0.861.

Conclusion

This proof-of-concept study shows that H&E pre-operative biopsies from TNBC, by means of deep learning technology, contain valuable information having predictive value for the outcome of NAC resulting in an AUC value of 0.696 outperforming a predictive AUC value of 0.63 based on structured clinical data of histological tumor grade, TILs, and ki-67 known from the literature.