Denoising for Intracranial Hemorrhage Images Using Autoencoder Based on CNN

Intracranial Hemorrhage, known as ICH, is a serious malady and is a major cause of stroke, disability, and death. Due to the malignant impact it would cause, it is essential to diagnose ICH in an efficient way. Meanwhile, the prevalence of artificial intelligence has caught people's attention and we considered that the diagnosis of ICH could be improved with the help of machine learning methods since these methods have succeeded in realms like speech recognition, visual object recognition, object detection etc. There have already been some computer aided diagnosis systems, based on machine learning methods, for helping the diagnosis of ICH, but most of them were used to classify the images and people seem to have ignored the importance of the quality of images. In this study, we applied a denoising method based on CNNs to improve the quality of ICH images based on CT. LeNet-5 and AlexNet were used due to their mighty ability to extract features. Autoencoder was used to reduce the dimensionality of the datasets to reduce the complexity of the data. Before denoising, we did some preprocessing in order to make our training faster by setting all images into a specific format. To test our model, three kinds of noise were added to the images, which were Salt&Pepper, Poisson and Gaussian. Lastly, we used Euclidean distance, which is the D-value of the pixel value between the original images and the processed ones, to measure the results of denoising images and determine their similarity to the original images. And the smaller the value, the better the effect. The results turned out that our denoising model is effective for all three kinds of noise, and the result of Salt&Pepper noise is the best.