Study On Low Dose CT Reconstruction Method Based On Deep Neural Network

Computed Tomography(CT)is widely used in medical practice as a diagnostic aid,but the high dose of radiation used in the process can cause serious harm to the body of patient,increasing the risk of cancer.Low-dose CT can be achieved by reducing the radiation dose or the number of projections.However,the projection noise obtained by low-dose radiation at each Angle is large,resulting in large noise in the reconstructed images,which affects the subsequent diagnosis.Removing noise in low quality images is an effective way of high-quality low dose CT reconstruction.Although the mathematical methods of image denoising have achieved great success,when they are used in CT denoising,the denoising performance is not good because of the uncertain noise type and non-uniform noise distribution in CT images.Deep learning,especially Convolutional Neural Network(CNN),is particularly suitable for image processing due to its powerful nonlinear fitting ability.Image denoising by deep learning is expected to achieve more accurate denoising effect than mathematical methods.The main work of this paper is as follows:(1)The DnCNN model based on channel domain attention mechanism is proposed to achieve the denoising of low-quality CT images.By constructing most of the convolution units into convolution units that integrate the channel attention mechanism,the important channel information is emphasized,the network performance is improved,and the denoising effect is improved.(2)The two-channel small UNet cascading model is proposed to further improve the denoising accuracy.The input images are processed separately through two channels to ensure the diversity of the generated feature images,which can better suppress noise.Simultaneously,residual learning is added to alleviate the gradient disappearance which may improve the network performance to a certain extent.(3)On the basis of algorithm research,a low-dose CT denoising system based on two-channel small UNet cascaded model is designed,which includes image uploading,the projection generation,the projection noise adding,the projection image with noise reconstruction and image denoising.Image denoising includes deep learning denoising and TV(Total Variation)model denoising.To suppress the noise of low-dose CT images,two denoising models of CT images were proposed in this paper,and their reconstruction performance was systematically evaluated and compared.Algorithm-based research,a low-dose CT denoising system based on two-channel small UNet cascade network is designed.These studies and designs have certain theoretical significance and practical value for low-dose CT reconstruction.