Application Research Of Image Segmentation Algorithm Based On Collaborative Attention Structure Fusion

In recent years,deep learning techniques received more and more attention in the field of image segmentation.Compared with traditional segmentation methods,it can extract deep semantic features that cannot be perceived by traditional methods,resulting in higher accuracy and generalization of the final segmentation effect.Therefore,this technology is widely used in the field of medical image processing,where skin lesion detection has been a hot topic for scholars.In the process of skin cancer detection,segmentation of skin lesion regions has always been a crucial and challenging step.So far,deep learning segmentation networks applied to segmentation of skin lesion regions are already in use and have achieved certain achievements.At the same time,it is still difficult to achieve accurate and clear segmentation.The original images may include unclear boundaries,varying sizes and shapes of lesion areas,as well as the presence of distractions such as hair,tattoos,and moles in some images,and the limited number of datasets to date,all of which can affect the final segmentation results.Currently,segmentation methods based on deep learning generally only use skip-connections to combine the feature maps of the encoder and corresponding decoder to recover the detailed information,and then use a simple convolution method to obtain the final feature map.Although this approach is easy to implement,the feature map cannot fully store the feature information of the encoding stage from the skip-connections,and the semantic feature information will be lost during the upsampling process of the feature map,resulting in poor generalization of the segmentation results.The feature map obtained through simple convolution method is also difficult to establish the relationship between features,affecting the segmentation quality.To solve the problem,this paper proposes a collaborative attention structure model based on encoder-decoder architecture.By building multiple modules based on feature fusion techniques,attention mechanism techniques,and multi-scale feature extraction techniques that assist the network to achieve finer segmentation together,the main work of this paper is as follows:(1)To address the problems of varying shapes of skin lesion regions,the difficulty of discerning the boundaries of lesion regions,and the presence of interferents as well as limitations of current segmentation methods,this study proposes a new skin lesion segmentation model to improve the representation capability of the network.First,in this paper,a new and improved feature extraction network MSCAN is used as an encoder for the shape characteristics of the dermoscopic lesion region.Second,in order to ensure efficient feature fusion between encoder and decoder and important feature perception during upsampling and skip-connection,this paper uses two attention modules with different ideas that can work together.The first one is a dense attention gates,which can be used to extract more features from external feature map including the encoder stage and the multi-stage decoder stages to improve the model’s ability to capture important features.The second one is a shuffle attention module,which includes improved channel attention and spatial attention modules to establish spatial and channel relationships of feature maps based on the internal feature maps,respectively,to realize the capture of detailed features and further improve the efficiency of feature map fusion.(2)To address the problem of limited number of images in datasets,this paper uses various data augmentation techniques to expand the datasets and help the model extract as many useful features as possible.To address the problem of uneven distribution of regions due to the varying sizes of lesions,a weighted loss function is used to help the model achieve better training,and the best weighting factor is determined experimentally.(3)In order to verify the reliability as well as the utility of the model,in this paper,a series of experiments are conducted on the ISIC-2017 dataset as well as on the PH2 dataset.This paper compares the proposed model with the segmentation model used for some skin lesions,and presents a series of ablation experiments.Finally,a comprehensive comparison between the model and the current classical segmentation network in terms of accuracy and parameter computation is made in this paper in order to prove that the network proposed can improve the accuracy of segmentation without introducing too many parameters and computation.It can determine the usefulness of the model in computer-aided diagnosis systems.