A deep learning approach for predicting outcomes of triple-negative breast cancer

Abstract: Breast cancer is the most common cancer in women. Triple-negative breast cancer affects 10-20% of breast cancer patients and is associated with an especially bad prognosis. Today, tissue slides are assessed manually by a clinician to set a prognosis. However, the prediction of outcomes could possibly be improved using machine learning. This work investigates various machine learning techniques for the task. In particular, a U-Net was trained on public data and was used to detect cells in microscopic images of H&E-stained triple-negative breast cancer tissue. The detected cells were then classified using logistic regression. The detected cells served as a basis for the prediction of local relapse, distant relapse, and overall fatality in a cohort of 155 patients diagnosed with triple-negative breast cancer. The accuracy of models fitted to features extracted using machine learning was compared to the accuracy of models fitted to features estimated by a clinician. The features extracted using machine learning was found to yield as good, or better, predictions compared to estimated features. A high number of tumor-infiltrating lymphocytes was associated with a better prognosis. This shows that machine learning can be used to find biomarkers in microscopic tissue images and use these to predict outcomes in cancer patients. The source code used in this work is published on Github: https://github.com/karvla/histosnet