Semantic Segmentation with DenseNets for Breast Tumor Detection

Breast cancer is one of the cancers with highest mortality rate, which are impairing the health of millions of women globally. Nowadays, the detection method of breast tumors based on breast ultrasound images (BUI) is semi-automatic. The localization of a region of interest (ROI) and the delineation of breast tumors within ROI need to be manually performed by an experienced doctor. In this work, we propose a fully automatic approach based on semantic segmentation to add contextual information in the complete image interpretation and avoid the ROI definition. The densely connected convolutional neural network is used for this approach to generate the feature maps. Dataset with 200 breast tumor images from Institute of Oncology in Warsaw was used to evaluate the performance of this approach. For DenseNet, the accuracy of our method is 99.2%. For traditional fully convolutional network, this accuracy is 85.60%. For the results of segmentation, the dice coefficient is equal to 0.8329. The results have demonstrated that the proposed method is accurate and objective for the segmentation of breast tumors in BUIs. Moreover, the robustness and generalization capacity of the method have been proven in this dataset.