关键词: 肺结节, 密集卷积, 跳跃连接, 编码路径, 结节分割

Abstract:

Many achievements have made in medical image segmentation based on deep learning networks. Because the round-shaped lung nodules are different from blood vessels and most of the lung structures are flat, a Bi-LSTM U-Net network with multi-view dense convolution is proposed to eliminate blood vessels by expanding U-Net and lung tissue structure to detect nodules. Unlike U-Net cascading in skip connections, the improved Bi-LSTM network combines the extracted feature map in the encoding path with the decoded convolutional layer for non-linear combination. In order to enhance feature propagation and encourage feature reuse, dense convolution is used in the last convolutional layer of the encoding path, and finally Batch Normalization（BN） is used to accelerate the convergence speed of the network. Experimental results show that the model effectively improves the accuracy of lung nodule segmentation, and the MIoU trained after extracting the axial position, coronal and sagittal views of each candidate sample in the LUNA16 and Alibaba Tianchi competition data sets reaches 90.1%.

Key words: pulmonary nodules, dense convolution, skip connection, coding path, nodule segmentation