RadImageNet

• Pneumonia, a lung infection causing inflammation, remains a leading cause of global mortality, resulting in an estimated annual fatality rate of approximately 4 million individuals according to the World Health Organization. Accurate and timely diagnosis of pneumonia in its early stages is crucial for effective patient care. Given the criticality of this issue, our study introduces an automated approach utilizing convolutional neural networks and transfer learning to detect pneumonia in chest X-ray images of children aged 1 to 5 years. We employed two pretrained ResNet-50 models trained on different datasets: ImageNet, containing 14 million natural images, and RadImageNet, consisting of 1.4 million medical images. Our results show that RadImageNet outperformed ImageNet, demonstrating superior performance in pneumonia detection. Evaluation using six key performance metrics revealed that the ResNet-50 model pretrained on RadImageNet achieved superior performance compared to the ImageNet-based model, further emphasizing the efficacy and interpretability of RadImageNet for binary classification tasks in medical image datasets. These findings underscore the importance of RadImageNet as a valuable source of pretrained models, particularly for small medical image datasets.

• In our approach, we utilized pretrained ResNet-50 models that were originally trained on ImageNet and RadImageNet, respectively. To tailor these models to our specific task, we removed the original output layer. This involved discarding the final classification layer, which was initially designed to classify images into numerous categories, and adapting it for our binary classification problem.

• Next, we froze all the layers in the networks, which means that we prevented their weights from being updated during the training process. This approach is commonly referred to as transfer learning, as we leveraged the knowledge and representations learned by the models on the respective datasets (ImageNet and RadImageNet) to benefit our own classification task.

• To customize the network for our specific task, we added a $\texttt{flatten layer}$ after the frozen layers. This layer reshapes the output from the previous layers into a one-dimensional vector, which serves as the input for subsequent layers. In addition, we included a dense layer with a $\texttt{sigmoid}$ activation function. This dense layer has a number of neurons equal to the desired number of classes in our binary classification problem. The $\texttt{sigmoid}$ activation function is commonly used in binary classification tasks as it outputs probabilities for the two classes or labels. The output of this layer represents the predicted probabilities for each class, which can be interpreted as the likelihood of an input image belonging to a certain class.

• Meanwhile, we utilized \textit{class activation maps} to facilitate the debugging of deep neural networks. To achieve this, we employed the Gradient-weighted Class Activation Mapping (Grad-CAM) algorithm for both models. By leveraging this technique, we gained valuable insights into the areas of interest within chest X-ray radiographs that the proposed network focused on while detecting pneumonia disease. The result of Grad-CAM is presented as a \textit{heatmap} visualization, allowing for a clearer understanding of the network's attention regions.