Two approaches for detecting pneumonia in x-ray images: description, implementation and comparison

The paper investigates two approaches for classification of x-ray images for presence of pneumonia. The first, widely used approach relies on neural networks (NN). The second approach utilises the theoretical concept of Kolmogorov complexity. The latter approach further enables the use of normalised compression distance (NCD) which defines a distance measure for arbitrary data objects, including images. Both approaches and their underlying algorithms are described in described in detail and implemented programmatically. The X-rays for this work are taken from the database of the Kaggle social network for data processing and machine learning. Optimal model parameters are found empirically. Experimental results show high accuracies for both approaches. This is a clear indication of reliability and applicability of either method for identifying the presence of pneumonia in x-ray images. The NN approach performs well when ample training data is available. The NCD method is turn applicable when training data is limited and the NN approach fails. These results provide a solid foundation for developing precise and reliable diagnostics of pneumonia, using a combination of the two approaches.