Predicting breast cancer recurrence and metastasis risk by integrating color and texture features of histopathological images and machine learning technologies

About 30%–40% breast cancer patients suffer from recurrence and metastasis, even after targeted therapy like trastuzumab. Since breast cancer recurrence and metastasis are intrinsically related to mortality, it is critical to predict the recurrence and metastasis risk of an individual patient, which is essential for adjuvant therapy and early intervention. In this study, we developed a novel breast cancer recurrence and metastasis risk assessment framework from histopathological images using image features and machine learning technologies. The detection framework was applied on a manually collected clinical dataset from the Cancer Hospital, Chinese Academy of Medical Sciences, consisting of 127 breast cancer patients with known prognostic information; and further independently validated on 88 formalin-fixed, paraffin-embedded (FFPE) samples downloaded from The Cancer Genome Atlas (TCGA) with known recurrence and metastasis status. As a result, the XGBoost-based method performed well using only 8 texture and color features, obtained internal testing AUC of 0.75 on clinical data and external testing AUC of 0.72 on TCGA FFPE data, respectively. In addition, this study found two important potential predictors, i.e., the second moment of the B color component and the detail level mean square error of the wavelet multi-sub-bands co-occurrence matrix. Our study benchmarked the performances of histopathological image features and machine learning technologies in the recurrence and metastasis risk assessment, and holds promise for relieving pathologists' workload and boosting the survival chances of the breast cancer patients.