Optimizing Pneumonia Detection from Scarce Chest X-Ray Data: A Comparative Analysis of Advanced Augmentation Techniques Using Deep Learning

Pneumonia, a prevalent respiratory ailment, necessitates precise and efficient diagnosis for effective treatment. This paper introduces a deep learning approach for pneumonia detection from chest X-ray images, employing two distinct network architectures: a convolutional neural network (CNN) trained from scratch and a pre-trained ResNet-50 model. The primary objective of this study is to conduct a comprehensive comparative analysis of advanced augmentation techniques, namely RandAugment, CutMix, MixUp, and Geometric augmentation, specifically designed to address the scarcity of data. To emulate scarce data conditions, the dataset was artificially down sampled. The experimental results yield significant insights into the impact of augmentation techniques on model performance. When utilizing the pre-trained ResNet-50 model, RandAugment exhibits superior efficacy compared to other augmentation methods, yielding an impressive accuracy of 90%. In contrast, when experimentation was conducted on the smaller CNN trained from scratch, geometric augmentation emerges as the optimal choice, achieving an accuracy of 83%. This comparative analysis underscores the critical importance of selecting appropriate augmentation techniques when confronted with limited data for pneumonia detection. The findings contribute to the optimization of deep learning models in the realm of medical imaging tasks, facilitating accurate and timely pneumonia diagnosis from chest X-ray images.