Diagnosis of Pneumonia Using Distribution of Pertinent Microstructures: An Application of Machine Learning

Detection of patients with pneumonia is an immensely important task that requires expertise and hence may prove costly. The automated methods of Chest X-ray analysis, especially pneumonia have gained attention in the last few years. This reduces cost and time and can be used for better visualization. In recent years the image processing community has worked on this problem with varying rates of success. This paper takes the work forward and applies Chest X-ray imaging to the diagnosis of pneumonia using the proposed Machine Learning pipeline, which involves pre-processing, extracting features, and selecting pertinent features. The feature extraction has been done using the Local Binary Pattern, and feature selection using the Fisher Discriminant Ratio. The proposed model has lesser memory and space complexity as compared to the often-used Convolutional Neural Network models which are Deep Learning models which do not require explicit extraction and selection. Furthermore, our model can handle imbalanced data gracefully. The results are encouraging and the proposed model may prove useful for medical eternity. Our model gives 87.3 % accuracy, 87.4 % sensitivity, and 87.1 % specificity. This model prepares the ground for Machine Learning to be used in the diagnosis of pneumonia. We are also working on another work that combines Conventional Machine Learning with Transfer Learning to accomplish this task.