| These years,deep learning has been normally used in intelligent medical image processing,medical record management,expert systems,and other "intelligent + medical" applications.In the field of intelligent medical image processing,semantic segmentation plays a very crucial role in many application scenarios.For example,formulate diagnosis and treatment plans,the first diagnosis of the disease,the exploration of the cause,etc.The semantic segmentation of retinal vessels is a crucial reference in the diagnosis of many ophthalmic and cardiovascular diseases and is a current research hotspot.However,the deep learning-based semantic segmentation model of retinal blood vessels generally has the problem of treating all the learned features equally,resulting in the background features with a large proportion being dominant in the learned features,the importance of blood vessel features is masked,and the model sensitivity and prediction accuracy are low.In addition,the deep learning models tend to be designed more and more complex,and simplifying the model scale while ensuring the model performance is also a difficulty in the field.For this reason,this article conducts indepth research on the semantic segmentation of retinal vessels based on deep learning on the aforementioned problems,with specific studies including:(1)A hybrid attention module based on supplementary signal g is designed to provide more accurate channel attention and spatial attention by adding supplementary signals carrying high-level semantic information into the hybrid attention mechanism.This hybrid attention module achieves the goal of being usable without changing the convolutional neural network(CNN)structure.(2)A hybrid attention-based retinal vessel segmentation model is constructed.Using the hybrid attention module designed in this paper,assign corresponding weights according to the primary and secondary features,suppress useless responses,and focus on important features.Retinal vessel segmentation features are identified while localizing vessel locations and validated in the DRIVE dataset.The results show that the sensitivity and accuracy of retinal blood vessel segmentation are effectively improved.(3)A method of reducing model parameters is used,getting a hybrid attention-based retinal vessel segmentation model optimized by asymmetric convolutions.The asymmetric convolution module is introduced into the segmentation model proposed in this article,which improves the sensitivity of the convolution unit to the detection of thin blood vessels,increases the robustness of the model to rotational distortion,and saves lots of training parameters of the model.The performance of the model is verified on the public datasets DRIVE,STARE,and CHASE DB1,and compared with existing methods,the hybrid attention-based retinal vessel segmentation model optimized by asymmetric convolutions in this paper can more accurately segment vessels,especially for small ones,and achieved the goal of reducing the number of model parameters,and saved the computational cost and storage space. |
