Noise reduction methods for the stroke detection and classification from brain computed tomography images

Hemiplegia, popularly known as stroke or paralysis, is the death of cells in the region as a result of the inability to feed any part of the brain for various reasons. It is the second leading cause of death in the world. However, according to WHO, more than 15 million people are diagnosed with stroke every year, and 5 million people die due to stroke. Time is the most important factor in stroke. The time-dependent increase in the effect of stroke is critical for both the patient's health and the variety of treatments that can be applied. For this reason, the rapid detection of stroke is important for the treatment of the disease. Medical imaging methods; It has a very important place in the diagnosis of diseases and in the course of treatment. The physician, radiologist or decision maker detects the disease through radiological images and evaluates the course of the disease, possible treatments and approaches. Computed Tomography (CT) is one of the medical imaging methods used in the detection of stroke disease. As with all medical imaging methods, pollution/noises that impair the quality of the CT images occur. This noise on the CT image can sometimes mislead the radiologist or decision maker both in the diagnosis of the disease and in further examinations. For this reason, removing this noise from CT images has been the motivation for this project. In general, stroke is divided into two as occlusive (ischemic) and hemorrhagic (hemorrhagic). In this study, the effect of the noise removal process in the classification of stroke present/no stroke, by reducing the noise in the computerized tomography images and improving the medical images, was investigated in this study. The noise removal methods examined in the study; Median and Gaussian filters from spatial field filters, Wiener filter and Wavelet Transform from transform field filters. In the study, the data set shared in the competition organized by the North American Radiology Society (KARD) was used. Gaussian noise on the obtained data set was added at 1, 5, 10, 25, 50, 75 and 100 standard deviation levels and cleaned with the aforementioned noise removal methods. SSIM, PSNR and MSE metrics were used to measure noise removal performance. In the study, deep neural network models ResNet50, DenseNet121, InceptionV3 and AlexNet were used for classification of stroke. The weights of the models to be trained were selected in two different ways: ImageNet weights and random. Accuracy, Precision, Sensitivity, AUC and F1 Score were used to measure the performance of the classification models. Noise removal and model classification performances were evaluated through graphics, and the relationship between noise removal and model performances was revealed. In the results of study; The effect of noise and noise removal on the performance of deep neural network models used to classify CT images is presented. Noise distorts the CT image and complicates the information extraction of deep neural networks. The performance of deep neural networks has been negatively impacted and degraded by noise. The effect of noise has been reduced by noise removal methods and the noise removal process has increased the performance of the classifier models compared to the noisy structure. With this study, researchers are suggested to add noise and noise removal as a step in their studies. In order to optimize the model performances, the use of different noise removal approaches in the noise removal stage, the parameter optimization of the methods used, the use of deep neural network models with different approaches in the studies, and the testing of noise removal with these deep neural network models are other recommended topics.