Abstract

Info

Title:: An algorithm of mass segmentation in mammogram by using deep convolutional neural network based on sliding patch

Author(s):: LIANG Nan1; 2; ZHAO Zhenghui3; 4; ZHOU Yi5; WU Bo1; 2; LI Changbo5; YU Xin3; MA Siwei3; ZHANG Nan1; 2; 1. School of Biomedical Engineering, Capital Medical University, Beijing 100069, China 2. Beijings Key Laboratory of Fundamental Research on Biomechanics in Clinical Application, Capital Medical University, Beijing 100069, China 3. Institute of Digital Media, Peking University, Beijing 100871, China 4. Laboratory of Mathematics and Applied Mathematics, Peking University, Beijing 100871, China 5. Department of Radiology, Huaihe Hospital, Institute of Medical Imaging of Henan University, Kaifeng 475000, China

Keywords:: Keywords: mammogram images breast mass sliding patch deep convolutional neural network image segmentation

Abstract:: Abstract: Objective An algorithm which includes local patch classification and breast mass segmentation in whole images was proposed based on sliding patch by using deep convolutional neural networks (CNNs) to provide effective morphological features for clinical diagnosis. Methods Firstly, breast region was extracted by regional growing algorithm and dilation algorithm, and the data were normalized. In order to obtain the diagnostic information of each pixel, the images blocks of mass patches and non-mass patches were slid and extracted in corresponding location of the original image. Based on the texture features extracted by deep CNNs, image blocks were classified. At last, based on the prospective classification results of the image blocks, the mass segmentation was made based on coarse-to-fine, and the pixel-level segmentation in whole image was obtained. Results Compared with the advanced deep CNNs, the experimental results demonstrated the algorithm achieved the best accuracy of 96.71% for patches classification under the model of DenseNet and the best F1-score of 83.49% for image segmentation in whole mammogram image. Conclusion According to the results achieved by CNNs, the proposed algorithm can segment mass in mammogram images with good generalization and robustness performance. And it provides a reliable basis for subsequent computer-aided diagnosis of breast lesions.