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。方法 对传统卷积神经网络框架进行改进，根据视盘的结构特点设计新的卷积融合层，能够更好地提升分割性能。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，本文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 本文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，只使用图像级标注就能够得到像素级标注的分割精度。缓解了医学图像中像素级标注难度大的问题。

Objective Ocular fundus image processing is one of the most popular research fields that combine medical science and computer science. Fundus images have the advantages of clear imaging, simple operation, and high efficiency, thereby enabling people to find various eye diseases as soon as possible. At present, deep learning methods provide state-of-the-art results on many tasks of image processing, including medical image segmentation and instance segmentation. A small number of objects are found in many cases of biomedical applications. Moreover, few datasets can be used. In most cases, fundus tests require a doctor to locate the optic disc and find its boundary. Therefore, retinal optic disc segmentation is an important problem in fundus image research. The success of the fully supervised learning algorithm relies on many high-quality manual comments/tags, which are often time consuming and costly to obtain. Different experts use different criteria, thereby resulting in some difficulties in medical image segmentation. If experiments with inaccurate data are conducted, not only will incorrect results be obtained but time will also be wasted. To save cost, this study proposes a constrained weakly supervised optic disc segmentation algorithm. Method By referring to the literature, we combine the convolution neural network (CNN) and the weak supervision method. A weak supervised learning method for sub-ocular image segmentation is proposed. First, the proposed visual CNN is pre-trained on a large auxiliary dataset, which contains approximately 1.2 million labeled training images of 1 000 classes. We can use this pre-training model to complete our own segmentation. Notably, we only use the parameters of the first five layers of the model to train our own models. Then, the top layer of the deep CNN is trained from RIM-ONE dataset. We fuse the conv3, conv4, and conv8 layers in our new model to improve the optic segmentation performance. Finally, we design a new constrained weak loss function to achieve an optimal output. The proposed loss function can optimize convolutional networks with arbitrary linear constraints on the structured output space of pixel labels. The key contribution of this study is to model a distribution over latent "pixel-wise" labels while keeping the network's output the same as the distribution. In this way, the output size is 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 methods are used to optimize the model. Result The proposed algorithm for constrained weakly supervised optic disc segmentation is evaluated with the RIM-ONE dataset. This method can effectively segment the contour of the video disc. The central part of the optic disc covered by blood vessels is well segmented. Our approach is evaluated in terms of mean accuracy, mean precision, and mean intersection over union. These three indexes are the common evaluation indexes in the field of image segmentation. We calculate the results prior to convolutional layer fusion and after convolutional layer fusion. Obviously, the latter results are better than the former ones. The latter results show that the mean accuracy in this work can reach 0.852, the mean precision can reach 0.831, and the mean intersection over union can reach 0.827; these findings are close to 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 uses orders of magnitude without annotation. With the model trained on the server, each image takes only a few seconds to predict. This prediction is faster than that of the method in the same type of some weakly supervised segmentation articles. Conclusion A new method to segment optic discs is proposed, and an end-to-end framework under deep weak supervision for image-to-image segmentation for medical images is developed. To preferably learn video disc information, we develop deep weak supervision for our formulation. Size constraints are also introduced naturally to seek for additional weakly supervised information. This work is the first to use image-level tags to conduct optic disc segmentation. The proposed models obtain more competitive results than the fully supervised method does. Experiments demonstrate that our methods achieve state-of-the-art results on weakly supervised medical images. The results can be applied to a wide range of medical imaging and computer vision applications. The research area on weakly supervised medical image processing has a broad prospect. An increasing number of people are expected to prefer the weak supervision method over the fully supervised method; even unsupervised learning is likely to cause a boom among scholars. These options can improve work efficiency and reduce labor costs. Experimental results also prove the effectiveness of our weakly supervised optic disc segmentation method.