Medical Image Denoising Method Based On Gabor Filtering And Deep Learning

In recent years,modern diagnostic medical technology has developed rapidly.Computed tomography(CT)has become an important tool because of its fast detection speed and low cost.It is often used to help diagnose complex fractures,tumors,etc.,and also plays a pivotal role in the fight against COVID-19(Corona Virus Disease 2019).As people pay more and more attention to health,it is generally hoped that CT radiation can be reduced.However,low-dose CT will cause image quality decline,resulting in misdiagnosis and other more serious consequences.For low-dose CT image denoising,according to the particularity of medical images,the shape and position of lesions in each patient are different,which means that detection depends not only on local features,but also on the overall global characteristics.The traditional method will result in blurred edges,unclear key points,lack of details and other problems.Due to its powerful feature representation capability,deep learning has gradually replaced traditional methods and become the mainstream direction of research.The main research contents of this paper are as follows:(1)The Bayesian Variational inference method is proposed to deduce Gabor filtering parameters.Experimental comparison shows that Gabor filtering is more suitable for the study of medical images than traditional CNN.Gabor filtering is used to replace traditional convolution to enhance the ability of edge detection.This method has less learnable weights and is less prone to overfitting,so the calculation operation is reduced significantly and the calculation speed is greatly improved.(2)This paper proposed a medical image denoising model BGFormer based on Transformer +Gabor,which combines the codec network and hop layer connection structure constructed by Le Win Transform module.It can significantly reduce the computational complexity of high resolution feature images.BGFormer has improved its ability to combine local and global dependency and boundary feature extraction in denoising tasks.We normalized the pixel values of the preprocessed image into the encoder.In the coding stage,the input of Le Win Transformer structure block was entered into bayes Gabor filter for convolution,and the encoded feature map was generated,and then the feature map was down-sampled.After two consecutive iterations of this process,the bottleneck layer is entered,where the encoded feature map is passed to another Le Win Transformer block that can be decoded at the same number of stages as it was encoded.After entering the decoder,at every stage of the decoder,and before sampling feature mapping on the connected by convolution,decoding phase can be seen as the opposite of the coding phase operation,and decoding the final figure after a block to produce the required output projection residual error,is obtained by subtracting the clear images of the final after removing noise.(3)The mixed variation loss calculation is proposed to comprehensively evaluate the calculation results from different angles,and improve the ability of the model to generate clear images by combining the changes of pixel distance,adjacent pixels and multi-scale perception.Compared with the previous deep learning-based denoising algorithm for the direction of medical images,the proposed method can recover clearer and clearer clean images with clearer structure,and shows more advanced performance in the peak signal-to-noise ratio(PSNR),root mean square error(RMSE)and structural similarity(SSIM)indexes.