Performance Evaluation and Comparison of Transfer Learning Models in Chest X-Ray Image Classification Using Deep Neural Networks

Detection of pulmonary diseases from X-ray images is one of the most significant challenges in medicine and deep learning. Traditional image analysis techniques are not efficient enough due to their reliance on manual features and functional limitations. Recent years have seen tremendous advancements in this area in deep neural networks (DNNs) and transfer learning models. This study aims to compare the performance of three popular chest (lung) X-ray (CXR) image classification models: ResNet50, MobileNetV2, and VGG16. Therefore, a dataset containing CXR images was initially prepared and preprocessed (normalized) by ImageDataGenerator. The dataset was then split into two sets: training (80%) and validation (20%). Then, the abovementioned transfer learning models were individually implemented and trained using this data[set]. The model performance was evaluated based on the following criteria: accuracy, confusion matrix, and classification report(s). The experimental results indicated all three models had acceptable image classification performance. ResNet50 exhibited higher accuracy in the validation dataset and consequently outperformed the other models. Also, MobileNetV2 was a suitable option for real-time applications due to its higher speed and smaller volume. On the contrary, VGG16 showed lower accuracy due to its older structure and lower complexity. Based on the results, the pulmonary disease diagnosis process could be effectively accelerated and its accuracy could be increased by adopting transfer learning models. Future research is recommended to employ hybrid models and more modern techniques like Transformers to enhance the results.