Breast Cancer Classification Based on Fully-Connected Layer First Convolutional Neural Networks

Abstract

Both Wisconsin diagnostic breast cancer (WDBC) database and the Wisconsin breast cancer database (WBCD) are structured datasets described by cytological features. In this paper, we were seeking to identify ways improve the classification performance for each of the datasets based on convolutional neural networks (CNN). However, CNN is designed for unstructured data, especially for image data, which has been proven to be successful in the field of image recognition. A typical CNN may not keep its performance for structured data. In order to take advantage of CNN to improve the classification performance for structured data, we proposed fully-connected layer first CNN (FCLF-CNN), in which the fully-connected layers are embedded before the first convolutional layer. We used the fully-connected layer as an encoder or an approximator to transfer raw samples into representations with more locality. In order to get a better performance, we trained four kinds of FCLF-CNNs and made an ensemble FCLF-CNN by integrating them. We then applied it to the WDBC and WBCD datasets and obtained the results by a fivefold cross validation. The results showed that the FCLF-CNN can achieve a better classification performance than pure multi-layer perceptrons and pure CNN for both datasets. The ensemble FCLF-CNN can achieve an accuracy of 99.28%, a sensitivity of 98.65%, and a specificity of 99.57% for WDBC, and an accuracy of 98.71%, a sensitivity of 97.60%, and a specificity of 99.43% for WBCD. The results for both datasets are competitive compared to the results of other research.