Abstract

Thisstudyanalyzestheperformanceoffourpre-trainedconvolutionalneuralnetworks(CNNs) (MobileNetV2, ResNet-50, DenseNet121andDenseNet201)inclassifyingchestradiographs to identify bacterial pneumonia, viral pneumonia or normal cases. MobileNetV2, the most efficient and accurate model, achieved an overall accuracy of 85%, excelling in the identifi cation of normal cases with 99% accuracy. This model was integrated into a Django-based websystem, whichallows physicians to upload radiographs, obtain automated diagnoses and visualize Grad-CAM heat maps for interpretation. The development followed the Scrum methodology, ensuring iterative progress and continuous improvement. The system aims to enhance diagnostic accuracy and accessibility, especially in resource-limited settings. Al though MobileNetV2 showed good results, its sensitivity for detecting bacterial pneumonia could be improved, suggesting that future improvements could be achieved with advanced data augmentation techniques and more extensive validation of the dataset. This work high lights the potential of lightweight CNNs in medical diagnostics and presents an efficient and scalable solution for early detection of pneumonia.

Statistics