The Automatic Segmentation of Mammographic Mass Using the End-To-End Convolutional Network Based on Dense-Prediction

Mammography is a common medical imaging using low-dose X-ray for the detection of breast cancer. The accurate discrimination in the mammogram between the normal issue and the mass is a critical step in the breast cancer diagnosis and treatment. However, it remains a great challenging task for physicians due to the variable morphological features of masses and the tanglesome biological structure. To solve this problem, we have proposed a convolutional encoder–decoder network which allows the input of the image of any-size and produce the segmented mask of the same size effectively from the mammographic image without any preprocessing. Consisting of full convolutional layers, each block in our model can update the parameters in the learning process. At the bottom of the encoder path, the receptive field has been broadened by using the multi-branch convolution operation instead of the single one. Additionally, the deep prediction blocks are introduced in the specific hidden layers to address the gradient vanishing problem during the back-propagation and speed up the convergence rate during the training phase. At the end of these blocks, the convolutional layer with 1 * 1 kernel is utilized to obtain the dense prediction and optimize the final result. The overhead bridge connecting the encoder and decoder components passes the low-level features in the encoder layers to the high-level ones in the decoder layers. We have evaluated our model on the public dataset DDSM and the private dataset from Zhongnan hospital of Wuhan University using several indexes including Accuracy, Precision, Specificity, Recall, and Dice coefficient. The experimental results demonstrate that the proposed method provides much better segmentation performance than several compared methods.