基于全卷积DenseNet的前列腺MRI图像分割新方法

前列腺磁共振（MRI）图像的自动分割对前列腺疾病的诊断至关重要，但是前列腺区域所占比例过小、组织边界模糊等问题为自动分割带来极大困难。针对这些问题，提出了一种基于全卷积DenseNet的前列腺MRI图像分割方法。该方法以现流行的深度学习理论为基础，利用迁移学习的思想，将DenseNet从自然图像迁移到前列腺数据集；采用反卷积和类似U-Net的全卷积神经网络结构，实现端到端的图像分割。同时引入并改进Dice相似性损失函数以解决前列腺MRI中背景所占比例远远大于前列腺区域和一些像素难以准确分割等问题。通过在PROMISE12数据集上进行实验，提出的方法Dice相似性系数达到93.25%，Hausdorff距离小于1.2 mm，相较于目前的主要方法，分割效果更好、所耗时间更短。

Novel prostate MRI segmentation method based on full convolution DenseNet

Automatic segmentation of prostate MRI images is very important for the diagnosis of prostate diseases. However, due to the small proportion of prostate regions and fuzzy tissue boundaries, automatic segmentation has become an important challenge. This paper proposed a novel prostate MRI segmentation method based on full convolution DenseNet. Based on the popular deep learning theory, the method applied the idea of transfer learning to load DenseNet from natural images to prostate datasets. The full convolution DenseNet used deconvolutions and the connection way similar to U-net, and achieved end to end image segmentation. Furthermore, improving the Dice similarity loss function could solve the problems that the proportion of background in the prostate MRI image is much larger than that of prostate region and some pixels are difficult to be accurately identify. The experiment on PROMISE12 dataset shows that the Dice similarity coefficient is over 93.25%, and Hausdorff distance is shorter than 1.2 mm. Compared with other main methods, the proposed method is more effective and takes less time.