Pediatric Pneumonia Diagnosis Using Cost-Sensitive Attention Models

Pediatric pneumonia is a medical condition in which air sacs of the lungs get filled with fluid. In recent years, chest X-rays have proved to be a better alternative to traditional diagnosis methods. Medical experts examine the chest X-ray images to detect the presence of pneumonia; however, the low radiation levels of X-rays in children have made the identification process more challenging leading to human-prone errors. The increasing use of computer-aided diagnosis in the medical field, especially deep learning architectures like Convolutional Neural Networks (CNNs) for images, helped tackle this issue. Our work proposes a Convolutional Block Attention Module (CBAM) attached to the end of pretrained ResNet152V2 and VGG19 with cost-sensitive learning. The weighted average ensemble uses weights which are calculated as a function of the precision, recall, f1-score, and AUC of each model. These values are concatenated as a vector and passed through a Tanh activation function. The sum of elements in this vector forms the weights. These weights when used in the weighted average classifier results in an accuracy of 96.79%, precision of 96.48%, recall of 98.46%, F1-score of 97.46%, and an AUC curve of 96.24% on the pediatric pneumonia dataset. The proposed architecture outperforms existing deep CNN models when trained with and without cost-sensitive training for the task at hand. We expect our proposed architecture to assist in real-time pediatric pneumonia diagnosis.