Research On Super-resolution Reconstruction Method Of Pathological Images Based On Dual Neural Networks

High-resolution pathology images are the objective criteria for high-precision diagnosis of diseases and are of great significance in the field of precision medicine.However,it is difficult to obtain high-resolution pathology images in real time,due to the limitations of hardware device resolution and scanning time of hardware device.Classical image super-resolution reconstruction algorithm can’t learn adaptively,and the training parameters of the image superresolution algorithm based on deep learning can’t be accurately estimated,resulting in the highresolution image generated being not realistic enough and the details are blurred,which is not suitable for pathological images super-resolution reconstruction.The paper focus on how to reduce network complexity and computational complexity while fully extracting and utilizing image feature information to reconstruct high-resolution pathological images.Therefore,two methods is proposed.(1)Previous medical image super-resolution algorithms mostly performed upsampling operations after extracting features in low resolution space,neglecting the many-to-one mapping relationship between low resolution pathological images and high-resolution pathological images,which cannot guarantee the quality of high-resolution pathological images.The paper proposes a lightweight sparse coding non-local attention dual network,which realizes the super-resolution reconstruction of pathological images by sparse coding non-local attention mechanism,Gaussian constraint and parameter sharing in the upsampling and downsampling dual branches.After the pathological image is reconstructed by this method,the peak signal to noise ratio reaches 30.84 d B,and the structural similarity reaches 0.914.The results show that this method can not only achieve the accurate reconstruction of highfrequency details in pathological images,but also effectively improve the efficiency of modeling by the lightweight sparse coding non-local attention mechanism.(2)The sparse coding non-local attention dual network not fully consider the conditional relationship between high-frequency information and low-frequency information of the image,the invertible transform block in the network cannot ensure the effective separation and fusion of low-frequency information and high-frequency information,and the hyperparameter selection in the network training process is set empirically.Therefore,based on the sparsecoding non-local attention stream model is proposed.The model adds a conditional invertible transform block to construct the conditional relationship between low-frequency information and high-frequency information,adds invertible 1×1 convolution to the invertible transform block to disrupt the channel order of the feature map to ensure the effective separation and fusion of high-frequency information and low-frequency information,and finally optimize the hyperparameters by Bayesian algorithm.Additionally,we introduce perceived loss and generative adversarial loss to ensure that the generated pathological image is highly consistent with the real pathological.The peak signal to noise ratio and structural similarity of the pathological images after this method reached 30.92 d B and 0.917.The results show that the sparse-coding non-local attention stream model can reconstruct high-resolution images more accurately.