Pneumonia X-ray Detection Based on DenseNet121 and Grad-CAM: Application of Explainable AI in Medical Imaging

This project builds an end-to-end deep learning pipeline that combines the DenseNet121 convolutional neural network with the Grad-CAM interpretability algorithm to achieve automatic screening of pneumonia in chest X-rays and visual diagnostic explanations. It aims to address the pain points of manual interpretation of medical images and enhance the credibility of AI decisions.

Pneumonia is one of the most impactful infectious diseases in global public health, consistently ranking among the top in terms of incidence and mortality. Chest X-rays are the first-line screening tool due to their low cost and wide availability, but manual interpretation is prone to fatigue and significant inter-doctor variability. Deep learning-based computer-aided diagnosis systems can provide consistent and scalable screening capabilities. This project aims to build an end-to-end pipeline that integrates explainable AI algorithms to enable automated, auditable, and clinically interpretable screening of pneumonia.

DenseNet121: Uses a dense connection mechanism where each layer receives feature maps from all previous layers as input and passes its own feature maps, effectively alleviating gradient vanishing, reducing parameter count, and promoting feature reuse.

Grad-CAM: Generates a localization map using gradients of the target class, highlighting regions with the highest decision weights, helping doctors associate AI-identified areas with pathological features and providing visual evidence support.

Data Processing: Inputs are uniformly resized to 224×224×3; batch size is set to 16; the training set applies augmentation strategies such as random rotation, dynamic scaling, and horizontal flipping.

Model Structure: Pre-trained DenseNet121 (ImageNet) + classification head (GlobalAveragePooling2D → Dropout(0.5) → Dense(128, ReLU) → Dropout(0.3) → Dense(1, Sigmoid)).

Training Strategy: Two-stage training—transfer learning (freeze the feature extractor and train the classification head) → adaptive fine-tuning (unfreeze the last 50 layers and optimize with a reduced learning rate).

Classification Report: Normal class precision: 0.94, recall: 0.85; pneumonia class precision: 0.91, recall:0.97; macro average:0.93, weighted average:0.92; ROC-AUC score:0.963.

Confusion Matrix: True negatives:199 cases, false positives:35 cases (technical artifacts), false negatives:11 cases (borderline/hidden cases), true positives:379 cases.

Grad-CAM generates heatmaps overlaid on original images; the hottest gradients are concentrated in pneumonia infiltration areas with correct anatomical focus. The network ignores external calibration marks, soft tissues, or text annotations, demonstrating strong robustness.

1. Multi-center external validation: Test the model's generalization ability on images from different medical institutions and devices;
2. Lung segmentation preprocessing: Combine a U-Net module to isolate lung regions and reduce noise interference;
3. Transformer comparative study: Develop Vision Transformers models and compare performance and attention map differences.

This project demonstrates that combining DenseNet121 with Grad-CAM provides a safe path for introducing computer-aided diagnosis systems in regulatory environments. It addresses the AI black box problem, generates visual activation maps, enhances transparency, auditability, and safety, helps doctors trust AI decisions, and meets regulatory review requirements.