Research On Retinal Fundus Image Segmentation Algorithm Based On Attention Mechanism And Multi-scale Features

The use of computer technology to achieve automatic segmentation of fundus images can better assist doctors in improving the accuracy and efficiency of diagnosing related diseases.Although some results have been achieved in vascular segmentation based on deep learning methods in recent years,there are still some shortcomings,such as the accuracy,sensitivity,and AUC indicators of segmentation that need to be improved.The segmentation quality of complex and micro blood vessels is poor,and the parameter quantity of the model is relatively high.In response to the above issues,this paper proposes a fundus image segmentation algorithm that combines attention mechanism and multiscale features.The main research contents and innovations of the paper include:(1)A novel fundus image segmentation model called AMMS-Net is proposed.Among them,a multiscale feature extraction and aggregation method based on Inception is designed to expand the receptive field of the model,aggregate the multiscale information of vascular features,and improve the recognition ability of the model for different forms of blood vessels;A pre-activation residual discarding module is proposed to alleviate the gradient disappearance problem of the model and accelerate the convergence speed of the model;A multiscale dense feature pyramid module based on ASPP is designed to achieve feature reuse of multi scale features;A residual spatial and channel attention module is proposed to achieve the information dependency between the vascular feature space and channels,and to improve the sensitivity and accuracy of model vascular segmentation.(2)Although AMMS-Net can achieve good segmentation performance,the segmentation quality for some tiny blood vessels is poor,and the model has a large number of parameters and high complexity.In order to achieve a lightweight and efficient fundus image segmentation algorithm,we propose the LSCA-Net network model.Among them,a structured convolution method for codec networks is designed to solve the limitations of single scale convolution operations and enhance the ability of the model to obtain feature context information;On the other hand,it improves the robustness of features and alleviates the problem of training over fitting of models;Adopt a dual attention mechanism based on space and channel to achieve feature recalibration and alleviate the impact of background and focus areas on segmentation quality;Use a lightweight design approach to optimize the complexity of the model and enhance its deployability in real-world applications.(3)We have trained and tested the algorithm models proposed in this paper on three datasets: DRIVE,CHASE＿DB1,and STARE.The experimental results show that compared with recent deep learning methods,the AUC value of AMMS-Net has achieved the highest value,reaching 98.57%,98.81%,and 98.67%,respectively.LSCA-Net realizes the lightweight of the model and further improves the performance of fundus image segmentation.The AUC values on the three datasets have reached 98.67%,99.05%,and99.02%,respectively.From the visual results,it effectively improves the segmentation ability of the model for complex and micro blood vessels.