Hyperparameter optimization of deep learning networks for classification of breast histopathology images

Abstract

After tumor detection in their breasts, women typically fear mastectomy; this affects curative care outcomes. Most tumors are benign. After resection and pathological examination, because of advances in medicine and treatment, the success rate of early breast cancer treatment can reach 60 to 90%. An accurate assessment of tumor extent is essential. In this study, a novel method of hyperparameter optimization of deep learning networks was proposed to classify tumors as malignant or benign. When setting hyperparameters in deep learning networks, most users use trial and error to determine them. In our experiments, the Taguchi method was used to select the impact factors. The orthogonal table design was used to conduct experiments. Then, the best combination of parameters was determined and significant impact factors were analyzed. The Breast Cancer Histopathological Database was used for analysis. This database was built in collaboration with the P&D Laboratory and contains 2480 benign and 5429 malignant samples. The experimental results showed that the proposed method obtained a high accuracy of 83.19%.