Research On Medical Image Segmentation Algorithm Based On Lightweight Attention Convolutional Neural Network

Medical image segmentation is a kind of task to segment interesting or diseased areas in medical images.At present,the method combining attention and convolutional network is mainly used,which can model local information and global information of image.However,this method still has some problems: 1)The training data of medical images are usually small,which leads to certain error deviation in attention calculation.The method is based on the combination of attention and convolutional network lacks error correction ability for attention calculation information,and attention information is lost in the downsampling.In addition,the calculation of attention will lead to the high cost of model calculation;and 2)Medical image data with annotated is generally scarce,and a lightweight and high-precision semi-supervised medical image segmentation method based on the combination of attention and convolutional network is lacking clinically.To solve the above problems,this paper conducted the following research:(1)For problem 1,this paper proposes a lightweight medical image segmentation method based on gated feature fusion.Firstly,a feature rearrangement module is proposed to solve the problem that attention information is lost in downsampling.Then,a gated feature fusion attention module is proposed.Axial attention is used internally to simulate the calculation of attention,which solves the problem of high cost of attention calculation.Moreover,the gated feature fusion module is used internally to correct the error deviation of attention calculation,which solves the problem that the model does not correct the attention information.(2)For problems 2,this paper first proposes a lightweight medical image segmentation method based on information interaction.Based on the lightweight medical image segmentation method based on gated feature fusion,a gated hybrid convolution module is proposed to replace the codec layer,and the calculation of attention is transferred to the skip connection part.An information interaction module is proposed to calculate and correct the attention of the codec layer,which solves the problems of high computing cost and no correction of attention information.Compared with the lightweight medical image segmentation method based on gated feature fusion,the computational cost index Flops decreased by 40%.Finally,in order to solve the problem of less annotation of medical data,this paper proposes a lightweight semi-supervised learning method based on the lightweight medical image segmentation method based on information interaction,which uses contrast loss to learn the intrinsic attributes of unlabeled data.Under the same labeled data,compared with the lightweight medical image segmentation method based on information interaction,the average intersection ratio of comprehensive performance indexes of the semi-supervised method on different data sets has been improved by more than 2.The method proposed in this paper is tested on four publicly available medical image segmentation datasets: ISIC 2018,Glas,Mo Nu Seg and TNBC.As for the average intersection ratio of comprehensive performance indicators,compared with other optimal methods,the lightweight medical image segmentation method based on information interaction proposed in this paper has an improvement of 1.62 percentage points in ISIC 2018,1.49 percentage points in Glas,3.69 percentage points in Mo Nu Seg,and 2.48 percentage points in TNBC.