A Convolutional Neural Network ensemble model for Pneumonia Detection using chest X-ray images

Pneumonia is a respiratory infection caused by microbes and other environmental factors. It infects the lungs causing a buildup of fluid and difficulty in breathing and is the leading cause for death in children under the age of 5 years. Timely detection proves essential in preventing adverse consequences including death. However, most areas in underdeveloped and developing nations do not have access to conventional diagnostic measures, preventive measures and adequate expert treatment. Computer-aided systems based on machine learning techniques can aid this task. However, most smart diagnostic systems may have the drawback of requiring extensive hardware and heavy computation power. The objective of this experiment is to develop a lightweight, deployable and accurate model to aid in the detection of Pneumonia. A Convolutional Neural Network architecture utilizing three different models of varying kernel sizes was developed. The outputs of these models were combined using a novel weighted ensemble approach which proposes an adjustable threshold value to change the model’s diagnostic capabilities as required. The flexible threshold value provides a means to adjust the weightage given to each model’s output and hence change the classification result depending on the actual case on hand. The model was evaluated on metrics including accuracy, recall, precision and f1-score and was able to achieve a high recall value of 99.23% with an f1-score of 88.56% which are critically high values for the given domain resulting in almost no chances of a Pneumonia positive case being misclassified. The absence of transfer learning or deep neural networks makes the model lightweight and hence, a plausibly deployable diagnostic-aid solution. Further studies were carried out to find methods such as – larger dataset, better preprocessing and more – to improve the model performance.