Diagnosing Chest X-Ray Images of Unseen Style Using Transfer Learning

Accurate and robust diagnostic tools are in critical need for the healthcare domain. This study proposes a novel method for diagnosing COVID-19 and pneumonia from chest X-ray images with an emphasis on the generalizability of the model across diverse image sources. Existing studies often rely on single-source datasets for training and evaluating machine learning algorithms, leading to limited generalization capabilities. This research highlights limitations of such studies and introduces a two-dataset approach, utilizing both a primary dataset for training and testing and a secondary dataset for rigorous out-of-domain testing, simulating real-world challenges. In this study multiple Convolutional Neural Network (CNN) based models were trained for the task. The model's performance was rigorously evaluated on two datasets: StyleA (collected from a single source which is used for both training and testing) and StyleX (collected from diverse sources which is used only for testing). Different image processing techniques and feature descriptors such as Contrast Limited Adaptive Histogram Equalization (CLAHE), Binary Image Conversion, and Histogram of Oriented Gradients (HOG) were applied to improve models generalization capabilities. Results show that applying these techniques improves models generalizing performance significantly.