Abstract

A novel combination of algorithms along with convolutional neural networks (CNNs) to denoise medical images is proposed in this paper. CNN is a type of deep learning model that specializes in retrieving information from input images instantly and capability to reduce the need for expert knowledge when extracting and selecting features. Hyper parameters like activation functions can have a direct impact on the model's performance in CNN and hence the paper, Improved Rectified Linear Units (I-ReLU)-CNNs are proposed for denoising the Medical Images(MI). In addition, the system also uses an encoder-decoder structure to save the most significant features of medical images while discarding the ones that are not required and by using a residual learning technique, the network is trained from start to end. Existing denoising techniques extract the clean image directly rather than learning the noise from the noisy image, where the denoised images are acquired through a proper learning method and the improved CNN layer. Furthermore, to optimize the weight ratio of CNN, different combinations of optimization algorithms are analysed and the network layer weight is optimized to get a denoised and clear image of MI. The proposed method not only denoises the MI but also completely removes the unwanted artifacts that would interfere with the blood cell counting stage. Extensive simulation is carried out and the noise is effectively removed and the mean square error (MSE) is reduced to zero. Thus the proposed optimized Denoising Convolutional Neural network (DnCNN) method for MI processing achieves excellent performance with reduced MSE and the desirable performance parameters.