Published Paper Details:

Pneumonia is an inflammatory condition of the lung primarily affecting the small air sacs known as alveoli caused by microorganism infection which can be viral or bacterial. It remains a critical public health concern worldwide, particularly in low resource settings, affecting severely specific age groups such as newborns & infants under 2 years old and adults above 65, due to their weak immune system. While Chest radiograph imaging is the most well-known screening approach used for detecting pneumonia in the early stages, its blurry and low illumination nature may call forth human error in manual diagnosis. Hence, the contribution of this work is the development of a two-stage pneumonia detection Expert System fusing the capabilities of both ensemble convolutional networks and the Transformer mechanism. In the first stage, a binary classification ensemble model is employed to detect whether a given chest X-ray indicates pneumonia or not. Upon a positive detection, the second stage activates a multi-class classification ensemble model that further categorizes the pneumonia into viral or bacterial, thus providing a finer level of diagnostic detail. The ensemble learning extracts strong features from the raw input X-ray images in two different scenarios: ensemble A (i.e., DenseNet201, Xception and InceptionResNet V2) and ensemble B (i.e., DenseNet201, Xception and VGG-16). The proposed ensemble deep learning model recorded 95.95% classification performance in terms of overall accuracy and F1-score for the binary classification task, while it achieved 88.57% for multi-classification task. To ensure that the diagnosis is not only automated but also interpretable to end-users, including healthcare professionals, the model is fed to an expert system where users can upload X-ray images, get a classification result and see the highlighted region which supports the diagnosis through Local Interpretable Model-Agnostic Explanations (LIME), a black box testing strategy. The proposed framework could provide promising and encouraging explainable identification performance compared to the individual or existing ensemble models building trust of users and healthcare professionals on the result.