Abstract:
Pneumonia is still a significant public health issue at worldwide level, especially in children and elderly, in which a not timely recognized or mistaken diagnosis may induce important and lethal complications. Pneumonia detection is a common application of chest X-ray image; however, it is time-consuming and subject to inconsistencies by the radiologists to manually interpret the results even with scarce resources in any given locale. In this thesis, a deep learning- based approach to automate the detection of pneumonia in chest X-ray images is proposed which leverages custom convolutional neural network (CNN) designs as well as transfer learning with state-of-the-art architectures such as VGG16, ResNet50, InceptionV3, and MobileNetV2. The dataset (of size 3392 X-ray images) was preprocessed and augmented for class imbalance and better generalization. Performance of models were evaluated using standard metrics such as accuracy, precision, recall and F1-score. Among all the architectures used, InceptionV3 performed best on the test set (accuracy 85.9%), It was well spread between pneumonia and normal classes. VGG16 worked fine and ResNet50 and MobileNetV2 overfit to the diminished training biased to the majority classes not even recognizing normal. These findings highlight the promise as well as challenges of transfer learning for improving diagnosis, including dataset imbalance and interpretability, as well as clinical workflow limitations. The results of this study contribute to the literature using AI in medicine and healthcare by demonstrating the feasibility of deep learning for pneumonia detection, laying a foundation for future work involving advanced clinical AI research.
