A comparison chart consisting of three axial chest CT scans illustrating radiological classifications of pneumonia. Panel (a) shows the consolidation type, characterized by areas of increased pulmonary density and patchy opacification within the lung parenchyma, consistent with alveolar filling that obscures underlying vascular structures. Panel (b) depicts the reticular type, showing a fine, net-like meshwork of linear opacities reflecting interstitial involvement or thickened interlobular septa. Panel (c) illustrates the nodular type, marked by multiple discrete, rounded opacities or small nodules scattered within the lung field. These images serve as educational reference for identifying long-term postoperative pneumonia (LTPP) patterns, emphasizing the morphological distinctions between alveolar consolidation, interstitial linear patterns, and nodular presentations. The imaging modality is high-resolution computerized tomography (CT), focused on the thoracic cavity and lung parenchyma for diagnostic classification in respiratory medicine.

This comparison chart displays a grid of Grad-CAM attention map heatmaps superimposed on pediatric chest X-rays, demonstrating model interpretability for pneumonia classification. The visual is organized into four columns representing three different deep learning architectures: Resnet50, SeResnet50, and a proposed 'Our model'. The rows are categorized by etiology: 'Bacteria' (top two rows) and 'Virus' (bottom two rows). For each model, two representative chest X-ray cases are shown per pathology. The heatmaps use a color gradient (red being highest intensity, blue lowest) to indicate which anatomical regions the AI prioritized for its diagnosis. In bacterial pneumonia cases, 'Our model' shows highly localized, intense 'hot spots' concentrated on pulmonary consolidations compared to the more diffuse attention of Resnet50. For viral cases, the heatmaps illustrate attention distributed across bilateral lung fields, reflecting typical interstitial patterns. This image serves as a diagnostic imaging tool for researchers and medical students to understand how computer vision models identify radiological features of pediatric pneumonia.

A multi-panel comparison chart illustrating various artificial intelligence (AI) interpretability techniques applied to Chest X-ray (CXR) images for pneumonia detection. The chart evaluates six local interpretability methods: Occlusion, Guided Backpropagation, Saliency, Integrated Gradients, Input X Gradient, and DeepLIFT. Each section includes a baseline 'CXR with Annotation' showing expert-identified pathology zones alongside heatmaps or saliency maps from five neural network architectures: DenseNet161, InceptionResNetV2, InceptionV3, ResNet18, and ResNet34. The heatmaps use color gradients (yellow to purple) to indicate regions of interest that influenced the model's classification. The visualizations demonstrate significant variability in model focus; for instance, the Occlusion method shows concentrated highlighting in the lower lung fields for ResNet models, while InceptionV3 exhibits more diffuse, high-intensity regions across the lungs. Other methods like DeepLIFT and Saliency display more granular, pixel-level importance. This comparison serves to evaluate the clinical validity of AI decision-making by comparing model-highlighted 'important' features against human-annotated radiological signs of pneumonia.