Effective pneumonia detection from chest x-ray images using Feature Extraction and Deep Transfer Learning with Stochastic Neighborhood Embedding

Abstract

Pneumonia is an infection that occurs in the lungs, posing significant global health challenges. It requires a particularly demanding treatment process for patients utilizing healthcare services. The application of artificial intelligence (AI) technologies in healthcare could address this issue. AI models supporting radiology experts can minimize factors like fatigue by reporting crucial details that might be overlooked in early diagnosis. These systems provide an automatic approach that does not require specialized expertise and produces rapid results.In this study, two different approaches are proposed for classifying lung X-ray images. The dataset used consists of X-ray images provided by the Indian Institute of Science, PES University, MS Ramaiah Institute of Technology, and Concordia University (DOI: 10.17632/9xkhgts2s6.3). This dataset comprises a total of 7927 images, with 3270 normal and 4657 pneumonia images, each of varying dimensions. In the first approach, features were extracted from the images using MobileNetV2, a pre-trained deep learning model. These features were then classified separately using Support Vector Machine (SVM) and Long Short-Term Memory (LSTM). SVM achieved an accuracy of 97.22% with an average prediction time of 8.2 milliseconds, while LSTM achieved an accuracy of 90.35% with an average prediction time of 9.78 seconds. In the second approach, features were extracted using MobileNetV2, and their dimensionality was reduced using t-Distributed Stochastic Neighbor Embedding (t-SNE) before being classified separately by SVM and LSTM. SVM achieved an accuracy of 94.64% with an average prediction time of 3.2 milliseconds, while LSTM achieved an accuracy of 94.83% with an average prediction time of 8.62 seconds. The t-SNE reduced the classification accuracy of SVM but increased the accuracy of LSTM, and it also shortened the prediction times for both classifiers. This study evaluates the performance of deep learning and machine learning models, as well as dimensionality reduction algorithms like t-SNE, in classifying lung X-ray images for pneumonia diagnosis. The findings suggest that computer-aided methods, particularly deep learning and dimensionality reduction algorithms, can effectively speed up and improve the accuracy of the diagnostic process for pneumonia. These results can serve as a significant guide for developing future pneumonia diagnosis methods.