Pathology image features as biomarkers for breast cancer diagnosis: A machine learning study

Introduction & Aim: Breast cancer is the most commonly diagnosed cancer with high mortality and morbidity. There are different biomarkers which be obtained from breast tissue or another biological fluids. The main aim of this study was to find how features extracted from breast pathology image can be used as diagnostic biomarkers.Methods: Pathology image data were retrospectively obtained from 530 breast cancer including 203 malignant and 327 benign cases diagnosed with fine needle aspiration (FNA). 30 different image features including intensity, texture and shape extracted and were used as input for machine learning methods. To find the best prediction model, five different machine learning methods were tested and their accuracy, area under the curve (AUC), precision, sensitivity and specificity, recall and f-measure calculated.Results: Among tested machine learning methods including K-nearest neighbors, Decision tree C4, Logistic regression, artificial neural network (ANN) and Naive Bayes, ANN had the best results. In this method, AUC = 0.99, accuracy = 0.96, sensitivity = 0.96 and specificity = 0.97 were obtained. Conclusion: Machine learning methods are powerful and easy tools for breast cancer diagnosis. Using these methods, malignant and benign breast cancers can be distinguished with highest accuracy and sensitivity. Our results not only demonstrated the application of the proposed approach on breast cancer diagnosis, but also physicians can benefit from saving the time and better understanding the properties of different types of breast cancer.