基于3D CNN的鼻咽癌CT图像分割

鼻咽癌CT图像分割是鼻咽癌诊断和治疗的先行任务,然而,由于鼻咽癌细胞的外形多样、灰度不均匀、边界模糊、病变形状复杂等因素使得分割难以准确。针对这一问题,提出了一种基于三维深度卷积神经网络的鼻咽癌CT图像分割方法,三维深度卷积神经网络框架的前5层采用卷积核为3~3的普通卷积,中间6层采用空洞率为2的膨胀卷积,后6层采用空洞率为4的膨胀卷积,每2个卷积层之间有一个残差连接,最后利用Softmax函数对每个像素点进行分类。膨胀卷积有助于得到精确的密集预测和沿物体边界的精细分割图,残差连接使深度卷积神经网络中的信息传播平滑,并能提高训练速度。实验结果表明,在鼻咽癌CT图像分割中该方法与其他主流方法相比有更好的性能。

A nasopharyngeal carcinoma CT image segmentation method based on 3D CNNs

Nasopharyngeal carcinoma computed tomography (CT) image segmentation is an essential task for diagnosis and treatments of nasopharyngeal carcinoma. However, nasopharyngeal carcinoma cells have various shapes, uneven gray scales, fuzzy boundaries, and complicated shapes of lesion cells, so it is difficult to accurately segment the image. In order to solve this problem, we propose a nasopharyngeal carcinoma CT image segmentation method based on three-dimensional convolutional neural networks (3D CNNs). In our three-dimensional deep convolutional neural network framework, ordinary convolutions with 33 convolution kernel are employed in the first 5 layers, the dilated convolutions with a dilation factor of 2 are employed in the middle 6 layers, and the dilated convolutions with adilation factor of 4 are employed in the last 6 layers. The residual connection is used between every two convolutional layers, and the softmax function is used to classify pixels. Dilated convolutions help to obtain accurate density prediction and fine segmentation maps along object boundaries. Residual connections smooth the information propagation in the deep convolutional neural network and improve the training speed. Experimental results show that the proposed method has better performance than other mainstream methods for nasopharyngeal carcinoma CT image segmentation.