ABSTRACT
Deep Learning (DL) is a popular machine learning algorithm, which proves its robustness on image recognition with high-classification performances. The DL increases the variety of applications due to data diversity and big data in recent years. One of the major superiorities of the DLs is a feature learning stage. Feature learning stage enables feeding the classifier with the responsible feature maps by extracting the low-, middle- and high-level features at specific depths instead of directly transferring the input. Although the DL algorithms have been frequently applied on object detection, face recognition, human pose recognition and semantic segmentation, its applications on medical image analysis are increasingly being studied. As a result of this, DL was applied to 2,000 chest X-ray images to perform organ segmentation. The chexNet radiography database was analyzed to train the convolutional neural network model. The study aims to segment the heart, right and left lungs. The DL-based organ segmentation is the first step of the analysis. The main contribution is clarifying the visibility of lung airways that are the most responsible and meaningful regions especially for obstructive respiratory diseases and increasing the clarity of the pattern, which is difficult to determine and interpret directly on radiography images, using local histogram equalization and adaptive thresholding methods exclusively on lung segments instead of the whole chest X-ray image. The proposed DL-based organ segmentation achieved a training accuracy rate of 99.75%. It achieved intersection-over-union rates of 96.32%, 98.67% and 94.89% for right lung, left lung and heart, respectively.
