Class-Based Attention Mechanism for Chest Radiograph Multi-Label Categorization

This work focuses on a new methodology for class-based attention, which is an extension to the more common image-based attention mechanism. The class-based attention mechanism learns a different attention mask for each class. This enables to simultaneously apply a different localization procedure for different pathologies in the same image, thus important for a multilabel categorization. We apply the method to detect and localize a set of pathologies in chest Radiographs. The proposed network architecture was evaluated on publicly available X-ray datasets and yielded improved classification results compared to standard image based attention.