Neural Network Based Classification of Breast Cancer Histopathological Image from Intraoperative Rapid Frozen Sections

Breast cancer is the leading cause of cancer-related mortality in women worldwide. Despite the rapid developments in diagnostic techniques and medical sciences, pathologic diagnosis is still recognized as the gold standard for disease diagnose. Pathologic diagnosis is a time-consuming task performed for pathologists, needing profound professional knowledge and long-term accumulated diagnostic experience. Therefore, the development of automatic and precise histopathological image classification is essential for medical diagnosis. In this study, an improved VGG network was used to classify the breast cancer histopathological image from intraoperative rapid frozen sections. We adopt a transformed loss function by adding a penalty to cross-entropy in our training stage, which improved the accuracy on test data by 4.39%. Laplacian-4 was used for the enhancement of images, which contributes to the improvement of the accuracy. The accuracy of the proposed model on training data and test data reached 88.70% and 82.27%, respectively, which outperforms the original model by 9.39% of accuracy in test data. The process time was less than 0.25 s per image on average. Meanwhile, the heat maps of predictions were given to show the evidential regions in histopathological images, which could drive improvements in the accuracy, speed, and clinical value of pathological diagnoses. In addition to helping with the actual diagnosis, this technology may be a benefit to pathologists, surgeons, and patients. It might prove to be a helpful tool for pathologists in the future.