Two Approaches for Detecting Pneumonia from Chest X-ray Images

Neural Network vs Kolmogorov Complexity

Abstract

Pneumonia is an infection that inflames the air sacs in the lungs. It remains the leading cause of death in children aged <5 years. This acute respiratory infection kills over 150,000 newborns yearly. We present two approaches for detecting pneumonic lungs. Both involve chest X-ray (CXR) image classification. The first approach is based on convolutional neural networks (CNN). The second approach, proposed by us, uses the theoretical notion of Kolmogorov complexity (KC), which introduces the normalized compression distance (NCD) – a way of measuring similarities between objects of different nature, such as images. The respective algorithms are described, software implementation details are presented. Experiments were conducted to enable us to choose optimal parameter values that would facilitate accurate pneumonia detection. The two procedures showed high classification quality. This convincingly indicates they were accurate in differentiating the chest X-rays. Though a known fact, the CNN approach was confirmed to be more efficient when dealing with a larger training dataset. On the other hand, the NCD-KC technique was shown to be more efficient when handling a small number of classified images. A more sensitive and more accurate pneumonia diagnosing technique that combines the strengths of both approaches is found to be feasible.