Abstract:

Chest X-ray (CXR) images are used to detect and differentiate diseases such as covid19, pneumonia, tuberculosis, and cancer. Many medical image analysis methods based on Deep Learning techniques have been proposed for health monitoring and diagnosis from CXR images. Deep Learning techniques have been used for various medical applications such as organ segmentation and cancer detection. Current studies in this area focus on the entire lung for disease diagnosis. Instead, it is thought that focusing on the left and right lung regions will improve the disease classification performance of Deep Learning algorithms. The aim of this study is to develop a model that will segment the lung region from CXR images using deep learning and image processing methods. For this purpose, a semantic segmentation model based on U-Net architecture, which is a deep learning method, has been developed. As it is widely known, U-Net shows high segmentation performance for various applications. U-Net is a different architecture composed of convolutional neural network layers and it gives more successful results in pixel-based image segmentation than classical models, even if there are few training images. For the training and testing of the model, the data set containing 138 chest X-ray images taken from the tuberculosis control program of the Montgomery County Health and Human Services Department, USA was used. The images in the dataset were randomly divided into 80% training, 10% validation, and 10% testing. The performance of the developed model was measured with the Dice Coefficient and the average value of 0.9763 Dice Coefficient was reached. It is important to crop the left and right lung regions detected by the model from the CXR images. For this, bitwise processing is applied to binary images with the image processing method. Thus, lung regions were obtained from CXR images. With these images, focusing on the cropped segment image instead of the overall CXR image can be used to classify many lung diseases.

Keywords: