Research On Fundus Image Segmentation Network Based On Domain Adaptation

Deep learning has been a hot topic in medicine in recent years.In some medically assisted diagnostic tasks,it has reached the accuracy comparable to manual segmentation by professional doctors.In CT/MRI medical images,the segmentation accuracy of some parts even exceeds that of doctors.The optic cup and optic disc of the eye can clearly reflect their physiological structure in the fundus picture.The ratio of the area of the optic cup to the optic disc can be used to determine whether the eye has certain diseases.This area ratio is called the cup disc ratio（CDR）.In healthy eyes,there is a fixed cup-disc ratio between the optic cup and optic disc,that is,the value of CDR is fixed.Conditions such as glaucoma and diabetes can make CDR abnormal.Therefore,accurate calculation of CDR can effectively screen some diseases of the eye.Nowadays,the technology of deep learning to segment optic cup and optic disc and calculate CDR has been very mature,and a lot of remarkable achievements have been achieved.Although the current methods based on deep learning have made numerous achievements,the following problems still exist in the segmentation task of optic cup and optic disc :（1）the existing methods rely too much on large-scale data sets,and it is known that the data sets of medical images are very scarce.（2）Optic disc itself has no clear physiological boundary,and the existing methods are generally not ideal for optic disc segmentation.In this paper,the segmentation method of semi-supervised optic cup and optic disc is studied deeply,and an effective semi-supervised method is proposed to reduce the dependence on large-scale data and data tags while ensuring the accuracy,and improve the generalization performance of network structure.Firstly,in order to get rid of the dependence on large-scale data sets and professional annotations,a dome-based adaptive method for optic cup optic disc segmentation network is proposed in this paper.Based on lightweight mobile-net-V2,it divides the training process into two parts: supervised learning with a small amount of data,and unsupervised adversarial learning with most of the other data using the trained network.In order to verify the effectiveness of the proposed network,sufficient experiments are carried out in this paper.Experimental results show that the proposed method achieves satisfactory results on Drishti and RIM-ONE data sets.Secondly,in order to improve the accuracy of semi-supervised network segmentation,attention mechanism is applied to optic cup optic disc segmentation task.A multi-scale spatial attention module is designed according to the feature that the attention mechanism can automatically focus on key areas.This module introduces a new Attention multi-fusion（AMF）module,which can obtain a larger receptive field and notice finer features compared to traditional spatial Attention mechanisms.Experimental results show that AMF greatly improves the accuracy of the model and reduces the accuracy loss caused by the semi-supervised method.Finally,in order to improve the generalization performance of network models,a metadata set enhancement module based on meta-learning is proposed.As the name implies,this module inherits the idea of learning generalization features quickly by meta-learning,and quickly makes the new data set fit the real features to improve the model generalization performance.The results of ablation experiment show that this module can improve the fitting speed and ability of the model.