Low-dose CT Image Denoising Based On Deep Learnin

Computed tomography(CT)is widely used for clinical screening,diagnosis,and intervention due to its fast scan time and clear images.However,the radiation during CT scans exposes patients to the risk of genetic damage and malignancy.Therefore,according to ALARA guidelines,low-dose CT(LDCT)examination is the future trend in the selection of clinical scanning modalities.But reducing radiation dose introduces quantum noise in CT,resulting in noise and artifacts in the reconstructed CT images,which may seriously interfere with the physician’s diagnosis of the condition.Traditional denoising methods are often based on the physical transformation of images and thus suffer from incomplete denoising or cumbersome computational processes.Deep learning techniques have been successfully applied to noise removal in LDCT images.Among them,Generative Adversarial Network(GAN)has very good image recovery and denoising ability.In this paper,we address the problem that the commonly used Mean Absolute Error(MAE)as the loss function causes the generated image details to be missing and smooth,and the problem of unstable training dynamics in GAN.The main works are listed as follows:(1)A loss function WP-MAE(Weighted-Patch MAE)for LDCT image denoising is proposed.The commonly used MAE loss function calculates the average loss per pixel,ignoring the differences in denoising difficulty in different regions of CT images.Therefore,it is improved on its basis by assigning suitable weights to the MAE losses of different regions of the image to balance their denoising difficulties.This loss function adjusts the model’s attention to different regions of CT images and improves the denoising effect of LDCT images.(2)A Reinforcement Learning-based Gradient Adaptive Generative Adversarial Network(RGA-GAN)model is proposed.While existing GAN models suffer from unstable training dynamics,the proposed model in this paper introduces multiple adversarial losses into the GAN-based denoising model to provide more gradient directions to enhance the stability of the model and uses reinforcement learning to assign weights to individual adversarial losses based on sample representations to prevent the model from being adversely affected by the unbalanced performance of individual adversarial losses.The effectiveness of the proposed framework is demonstrated on the testing dataset.In summary,after experiments,the proposed loss function and model in this paper demonstrate good results on the LDCT image denoising problem.