DeepSystem: The Effect of the Optimized Deep Learning and Blurring Filters on the Automated Detection of Pneumonia Using X-ray Images

Pneumonia is a potentially fatal respiratory infection affecting a significant portion of the population, particularly in areas with high pollution, overcrowding, poor sanitary conditions, and limited healthcare infrastructure. Pneumonia typically leads to pericardial effusion, a condition in which fluid fills the chest and causes breathing problems. Timely and accurate diagnosis of pneumonia is vital for effective treatment that improves the probabilities of survival. Specialists can detect pneumonia manually, but the process is time-consuming and prone to human error, making it inefficient for processing huge volumes of images. Automated detection systems for pneumonia can significantly streamline this process. This study investigates the power of deep learning to develop predictive models for accurate pneumonia detection using chest X-rays. It examines the impact of several factors on classification accuracy using ResNet-50 and Inception V3 as deep feature extraction models. These factors include the effect of applying four common image filters on classification accuracy, the influence of using a dropout layer, and the impact of employing different classifiers, i.e., Support Vector Machine (SVM), Random Forest (RF), and Naïve Bayes (NB). The findings reveal that, although the results across all models and filters were comparable, ResNet-50 combined with SVM scored the highest accuracy of 98% when using the Gaussian filter. Similarly, Inception V3 with SVM provided high classification accuracy, achieving 98% with both the Gaussian filter and the original data. However, the performance of the Median filter (using Skimage) showed improvement with Inception V3 compared to ResNet-50. These findings underscore the importance of selecting suitable image filters and deep learning models to optimize classification performance. Moreover, SVM consistently outperformed both RF and NB across all datasets, confirming its effectiveness as the most reliable classifier in this context.