摘 要: 目的 医学图像的像素级标注工作需要耗费大量的人力，针对这一问题，本文以医学图像中典型的眼底图像视盘分割为例，提出了一种带尺寸约束的弱监督眼底图像视盘分割算法，该算法只使用图像级标注就能够得到像素级标注的分割精度。方法 本文对传统卷积神经网络框架进行了改进，根据视盘的结构特点设计了新的卷积融合层，能够更好的提升分割性能；为了进一步提高视盘分割精度，本文对卷积神经网络的输出进行了尺寸约束，同时用一种新的损失函数对尺寸约束进行优化，所提的损失公式可以用标准随机梯度下降方法来优化。结果 在RIM-ONE视盘数据集上展开实验，并与经典的全监督视盘分割方法进行比较，实验结果表明，所提算法在只使用图像级标签的情况下，平均准确识别率(mAcc)，平均精度(mPre)和平均交并比(mIoU)分别能达到0.852，0.831，0.827。结论 所提的算法不需要专家进行像素级标注就能够实现视盘的准确分割，缓解了医学图像中像素级标注难度大的问题。
Algorithm for size constrained weakly supervised optic disc segmentation of fundus image
Abstract: Objective Ocular fundus image processing is one of the most popular research fields combining medical science and computer science. Fundus images have the advantages of clear imaging, simple operation and high efficiency, enabling people to discover various eye diseases as soon as possible. Nowadays, methods of deep learning provide the state of the art results on many task of image processing, including Medical images segmentation and in instance segmentation. A small number of objects is to be found in many case of biomedical applications. And there are few datasets can be used. Often fundus tests require a doctor to locate the optic disc and find its boundary. Therefore, retinal optic disc segmentation is a very important research problem in fundus images. The success of the fully supervised learning algorithm relies on a large number of high-quality manual comments/tags, which are often time-consuming and costly to obtain. Different experts have different making criteria, this will bring some difficulties to medical image segmentation. If you do experiments with inaccurate data, you will not only get the wrong results, but also waste time. In order to save the work cost, a constrained weakly supervised optic disc segmentation algorithm is proposed in this paper. Method By reading many literature, we combination with convolution neural network (CNN) and the weak supervision method, A weak supervised learning method for subocular image segmentation is proposed. Firstly, the proposed visual convolutional neural network is pre-trained on a large auxiliary dataset, which contains approximately 1.2 million labeled training images of 1000 classes. In this way, we can use this pre training model to complete our own segmentation. It is worth noting that we only use parameters of the first five layers of the model to train our own models. Then, the top layer of the deep convolutional neural network is trained from RIM-ONE dataset, and we fused the conv3, conv4, conv8 layer in our new model, this can improve the optic segmentation performance. At the end, we design a new constrained weak loss function to optimal output. The proposed loss function can optimize convolutional networks with arbitrary linear constraints on the structured output space of pixel labels. And the key of this paper is to model a distribution over latent ‘pixel-wise’ labels while the network’s output is as same as the distribute. This allows the output size to be within a reasonable range. The weak loss function is used to constrain the foreground and background sizes of the target. The KL divergence and Stochastic Gradient Descent were be used to optimize the model. Results The proposed algorithm for constrained weakly supervised optic disc segmentation is evaluated on the RIM-ONE dataset. This method can segment the contour of the video disc very well, and the central part of the optic disc which is covered by blood vessels is also well segmented. Our approach is evaluated based on mean accuracy, mean precision and mean intersection over union. The above three indexes are the commom evaluation indexes in the field of image segmentation. We calculated the results before convolutional layer fusion and after convolutional layer fusion separately. Obviously, the latter result is better than the former. The latter results show that the mean accuracy in this paper can reach 0.852, the mean precision can reach 0.831 and the mean intersection over union can reach 0.827, this is very close to the current state of the art result. We only use image-level tags without any pixel-level mask. Overall, our algorithm for constrained weakly supervised optic disc segmentation achieves 90% of the performance of the fully supervised approach, which using orders of magnitude less annotation work. By training the model on the server, each image takes only a few seconds to predict, which is faster than the method in the same type of article. Conclusion A new method to segment optic disc is proposed, an end to end framework under deep weak supervision to perform image to image segmentation for medical images are developed. To preferably learn the information of video disc, deep weak supervision is developed in our formulation. Size constraints are also introduced in a natural way to seek for additional weakly-supervised information. This is the first paper to use image-level tags to do optic disc segmentation, the proposed models obtain quite competitive results compared with the fully supervised method. Experiments demonstrate that our methods achieve state of art results on weakly supervised medical images. The paper also can be applied to a wide range of medical imaging and computer vision applications. The research area about weakly supervised medical image processing has a broad prospect. More and more people will turn from the fully supervision method to the weak supervision method, even unsupervised learning is likely to cause a boom among scholars. It can improve the working efficiency and reduces the labor costs. The experimental results also prove the effectiveness of our weakly supervised optic disc segmentation method.