Medical Image Denoising Using Convolutional Autoencoder with Residual Skip Connections in CT Images of Brain

Medical imaging plays a critical role in accurate diagnosis of diseases. Noise can be significantly affect the quality of images during acquisition that leads to potential misdiagnosis. It needs for repeat scans, which increases patient burden and healthcare costs. The dose of radiation in computed tomography (CT) directly impacts the quality of image and also has a significant health issue. It is essential to have low-dose radiation during scanning; however, it affects the image quality. Hence denoising methods based on deep learning have potential to be investigated to improve the quality of CT scans for image analysis. This paper proposes a deep learning architecture based on convolutional autoencoder(CAE) with residual skip connections for enhancing CT images. The autoencoder structure facilitates learning to represent a low-dimensional feature of the original image, while the residual skip connections ensure preservation of essential information details of the image. The model is evaluated against various noise types, including random, Gaussian and salt-and-pepper noise which are embedded in 2048 CT images collected from Kalinga Institute of Medical Science, Bhubaneswar, India. The effectiveness of noise removal methods in improving the quality of the image are evaluated. The best result was obtained with peak signal to noise ratio(PSNR) improvement of 51.6% and structural similarity index measure(SSIM) improvement of 170.3% as compared to noisy image. The results are also compared against the traditional denoising methods like Gaussian and Wiener filter implemented in this work, which shows better result in CAE based approach.