The ophthalmic images classification problem is one of the topics of computer-aided medicine,reasonably adding prior knowledge can significantly improve the performance of the classification model.A common way to add prior knowledge is to use a mask to guide the attention of the neural network.For example,in the disease classification task,we can weight the lesion masks to neural network feature maps in order to strengthen neural network attention to lesions.Generally,the manual annotation cost of lesion masks is relatively high due to the complex morphology characteristic of the lesions.The specific works of this dissertation are as follows:（1）In order to reduce the cost of annotation,this dissertation uses anatomical structure masks to replace lesion or image quality abnormal masks.In the process of using the mask to guide the neural network attention,the forward gradient activation mapping module is designed to obtain the attention area of the neural network.After that a new mechanism is proposed,the proposed mechanism uses dual thresholds to constrain the angle range between the "neural network attention region" vector and the "anatomical structure mask" vector,update the single-path gradients of vectors which don’t meet the angle range.In order to verify the effect of the above mechanism,this dissertation designs two verification tasks:Verification task Ⅰ: This dissertation designs a verification task that the "regional image quality" assessment of the bulbar conjunctiva and builds a large dataset about the task.This task utilizes the "regional image quality" assessment replaces the traditional image quality assessment in order to “accept”more images that meet the image quality.In the construction of the data set,this dissertation designs multiple auxiliary tasks to guide the judgment of the evaluation of the "regional image quality" of the bulbar conjunctiva.Experiments show that the auxiliary tasks can effectively improve the classification results of "regional image quality" of the bulbar conjunctiva compared with other classification networks.Verification task Ⅱ: This dissertation designs a retinal macular disease classification task based on two modal volume data.This task directly classifies volume data in order to strengthen the utilization of three-dimensional information,and combines two modal data to strengthen the neural network comprehension of lesion features.Moreover,this dissertation proposes a shallow feature fusion framework,in this framework several feature fusion modules are added to different positions of the neural network feature extractor in order to combine shallow features and strengthen non-linear expression of fusion features.In the feature fusion modules,a sub-structure is designed to strengthen abstract semantic expression ability of the shallow features.Experiments show that the feature fusion framework and modules proposed in this dissertation have a better classification effect than other feature fusion architectures on this task.After adding mask information,compared with other mask guidance methods,the mask guidance mechanism proposed in this dissertation can accurately guide the network to focus on a wide range of anatomical structure regions in the above two verification tasks,and have better classification effect.（2）A system is designed for the above work,which supports the image category labeling,classification model training,metrics analysis of the test images,and visualization of neural network attention region.