A Comparative Assessment of Transfer Learning Approaches in Pneumonia Detection

Artificial Intelligence in medical imaging is transforming healthcare, especially in critical areas such as pneumonia diagnosis. Effective treatment for pneumonia, a potentially fatal respiratory disease, depends on prompt and accurate diagnosis. The conventional manual interpretation of chest radiographs presents accuracy challenges and requires a substantial investment of time and resources. Deep CNNs provide effective automation, particularly when combined with transfer learning. This work assesses well-known pre-trained models for pneumonia classification (binary) using transfer learning. Using Google Colab to assess performance on a variety of X-ray datasets from Kaggle, VGG16 excels on small datasets (4000) and medium datasets (10,000), achieving an optimal F1 score and an accuracy of over 98%. RESNET50V2 achieves 99% accuracy on larger datasets (around 16,000). INCEPTIONV3 demonstrates adaptability while delivering competitive performance across datasets. This study contributes to an understanding of the model-dataset relationship in pneumonia detection by highlighting the impact of the model’s architecture, and the volume of the dataset on model performance. To improve AI-driven pneumonia diagnosis even more, future studies should investigate recent architectures and undertaking real-world validation studies to advance the field of study.