Automatic Classification System for Chest X-Rays Based on CNN and TensorFlow: A Deep Learning Project for Pneumonia Detection

This project is a medical image classification system published by keshavbs342 on GitHub (original link: https://github.com/keshavbs342/Project-Medical-Image-Classifier, published on May 31, 2026). Its core goal is to use Convolutional Neural Networks (CNN) and the TensorFlow framework to automatically identify pneumonia lesions in chest X-rays and provide AI-assisted support for medical diagnosis. The project addresses the efficiency and subjectivity issues of traditional diagnosis through deep learning technology, and has important clinical application value.

Pneumonia is a common and life-threatening respiratory disease, and early accurate diagnosis is crucial for treatment and prognosis. Traditional diagnosis relies on doctors' visual interpretation of chest X-rays, which is time-consuming and susceptible to subjective factors and fatigue. In recent years, breakthroughs in deep learning in the field of computer vision have revolutionized medical image analysis. CNN can automatically learn hierarchical features of images, perform well in image classification tasks, and assist doctors in quickly identifying pneumonia lesions, improving diagnosis efficiency and accuracy.

This project is a deep learning-based medical image classification system that实现 binary classification of chest X-rays into normal/pneumonia. Core functions include:

1. Image preprocessing: standardize original X-rays;
2. Feature extraction: automatically learn image features through CNN;
3. Binary classification prediction: output normal or pneumonia results;
4. Confidence evaluation: provide confidence scores for prediction results;
5. Visualization display: generate heatmaps of model focus areas.

Dataset

A public chest X-ray dataset containing normal and pneumonia categories is used. Preprocessing steps: unify size (e.g., 224x224), pixel normalization (0-1), data augmentation (rotation, flipping, scaling, brightness adjustment), and split into training/validation/test sets.

Model Architecture

Adopt a classic CNN structure:

• Convolutional layers: multi-layer structure to extract low/mid/high-level features, followed by batch normalization, ReLU activation, and max pooling;
• Fully connected layers: flatten feature maps and input to hidden layers, including Dropout to prevent overfitting, and the output layer uses Sigmoid to output binary classification probabilities. Optimization strategies: binary cross-entropy loss function, Adam optimizer, learning rate scheduling, and early stopping mechanism.

TensorFlow Implementation

Based on TensorFlow 2.x, use the Keras API to build the model, tf.data for efficient data loading, support GPU acceleration, and can export to SavedModel or HDF5 format.

Medical Image Preprocessing

• Size standardization: unify to model input size;
• Grayscale processing: ensure consistent channel count;
• Contrast enhancement: histogram equalization or CLAHE to highlight details;
• Normalization: normalize pixel values to 0-1 or Z-score standardization.

Data Augmentation

Geometric transformations (±15-degree rotation, horizontal flipping, 0.9-1.1x scaling), brightness adjustment, Gaussian noise injection (need to avoid medically unreasonable images).

Class Imbalance Handling

Class weights, oversampling minority classes, undersampling majority classes.

Model Interpretability

Grad-CAM heatmaps to show model focus areas; feature map visualization of learned features in middle layers.

Evaluation Metrics

Use multi-dimensional metrics such as accuracy, precision, recall, F1 score, AUC-ROC, and confusion matrix.

Test Set Performance

Typical results:

• Accuracy: 90-95%
• Precision: 85-92%
• Recall: 88-94%
• F1 score: 0.87-0.93
• AUC-ROC: 0.94-0.98

Note: Metrics are affected by dataset quality, distribution, and model architecture, and need to be validated with large-scale clinical data.

Application Scenarios

• Auxiliary diagnosis: primary screening, second opinion for doctors, prioritize high-risk patients;
• Telemedicine: remote expert support, 24-hour screening, reduce regional differences;
• Epidemic monitoring: handle surge demand, standardize processes, support public health decisions.

Limitations

• Data quality dependency: generalization ability is affected by equipment, shooting conditions, and population differences;
• Annotation accuracy: need multi-expert consensus to establish a gold standard;
• Lesion diversity: difficult to cover all pneumonia types;
• Confounding factors: easy to confuse with other lung diseases;
• Regulatory ethics: need to meet regulatory requirements such as FDA/NMPA, involving privacy and liability issues.

Future Directions

• Multi-class extension: normal/bacterial/viral pneumonia, etc.;
• Multi-modal fusion: combine clinical data with images;
• Lesion localization: pixel-level segmentation to quantify lesion range;
• Transfer learning: use ImageNet pre-trained models to accelerate training;
• Federated learning: integrate multi-center data under privacy protection.

Conclusion

This project demonstrates the potential of deep learning in medical image analysis, but AI is an auxiliary tool rather than a replacement for doctors. Its value lies in improving efficiency, reducing missed diagnoses, optimizing resource allocation, and ultimately improving patient prognosis. With technological progress and data accumulation, medical AI will play a more important role in precision medicine.