Lung CT Image Segmentation Using Reinforcement Learning.

Nowadays, lung disease is a major global problem in the world. Many of the lung diseases are untreatable and reduce the life span of the individual. Initial diagnosis of these diseases can help to completely cure the illness and reduce disease progression. Diagnosis of lung disease is often made based on CT scan images of the lungs. But the diagnosis of lung disease in CT images is a complicated and challenging task. Such a challenge is due to the noise of imaging, the lack of transparency of the image, the intensity of the lighting of some components, and so on. Eliminating these issues makes it easy to diagnose the disease for specialized physicians as well as existing automated diagnostic systems. In this paper, we propose a novel approach for lung CT image segmentation that solves some of these challenges. Also, correct accurate segmentation of these images can help to automatically detect lung diseases. The correct segmentation means a precise determination of the intensity of each area's brightness, the exact location of the pixels in each area and the preservation of the original image. We employed a reinforcement-based approach for our image segmentation. The proposed method produces the image without noise, and the image components are split with the best accuracy. To evaluate our approach, we manually generated 200 labelled lung CT images. After that, the proposed method, along with two recently published approaches is evaluated against our generated dataset. This comparative evaluation is conducted in two qualitative and quantitative ways. For the qualitative comparison, we employed the consultation of two ultrasound radiologists for evaluating the results of our approach. We also used a similarity measurement to quantitatively evaluate the performance of our proposed segmentation method. In both evaluation approaches, the proposed algorithm provides better performance compared to the other methods. The proposed method provides an accuracy of more than 90% for lung image segmentation, which shows around 5% increase in the accuracy comparing to the state of the art. The high accuracy of this method is due to its fitness to the application. [ABSTRACT FROM AUTHOR]