Representative chest X-ray images from the RSNA dataset, showing both normal cases and pneumonia-positive cases with annotated bounding boxes indicating areas of lung opacity.
Automated Pneumonia Detection using Deep Learning
Comparative Analysis of Faster R-CNN Architectures for Medical Image Classification
Project Overview: This project explores automated pneumonia detection in chest X-rays using three distinct Faster R-CNN implementations. We evaluated different backbone architectures, optimizers, and transfer learning approaches on the RSNA Pneumonia Detection Challenge dataset, achieving up to 77.34% training accuracy and 76.81% validation accuracy with ResNet-50 pretrained on ImageNet.
| Completion Date | December 2024 |
| Course Context | CS 5785: Applied Machine Learning Final Project |
| Dataset | RSNA Pneumonia Detection Challenge (26,000+ training images) |
| Key Technologies | PyTorch, Faster R-CNN, ResNet-50, Transfer Learning |
| Core Concepts | Object Detection, Medical Image Analysis, Deep Learning, Binary Classification |
Technical Approach
We implemented and compared three distinct approaches to automated pneumonia detection, each leveraging different aspects of the Faster R-CNN architecture for object detection and localization of lung opacities in chest radiographs.
Method 1: Torch Faster R-CNN with ResNet-50 and Adam Optimizer
Architecture: Faster R-CNN with ResNet-50 Feature Pyramid Network (FPN) backbone
Transfer Learning: Pretrained weights from COCO dataset
Training Configuration:
- Optimizer: Adam (learning rate: 0.001)
- Epochs: 2
- Batch size: 4
- Image resolution: 128×128 pixels
- Data split: 80% training, 20% validation
Method 2: Faster R-CNN with ResNet-50 (ImageNet Pretrained)
Architecture: ResNet-50 with custom classification head
Transfer Learning: Pretrained weights from ImageNet
Network Design:
- Feature extractor: ResNet-50 (frozen pretrained layers)
- Global Average Pooling layer
- Dense layer: 1024 units with ReLU activation
- Output layer: Sigmoid activation for binary classification
Training Configuration:
- Loss function: Binary cross-entropy
- Optimizer: Adam
- Epochs: 10 (early stopping at 1 epoch for optimal performance)
- Batch size: 32
Method 3: PyTorch Faster R-CNN with ResNet-50-FPN and SGD
Architecture: Faster R-CNN with ResNet-50-FPN backbone
Transfer Learning: PyTorch default pretrained weights
Training Configuration:
- Optimizer: Stochastic Gradient Descent (SGD)
- Learning rate: 0.01
- Weight decay: 0.0005
- Epochs: 2
- Data split: 80% training, 20% validation
Faster R-CNN Architecture
The Faster R-CNN architecture combines two key components for object detection:
Region Proposal Network (RPN): Generates object proposals by sliding a small network over the CNN feature map. For each location, it predicts object/no-object scores and bounding box coordinates.
Fast R-CNN Detection Network: Takes the proposed regions and performs classification and bounding box regression. The network outputs class probabilities and refined bounding box coordinates.
The loss function combines classification loss and localization loss:
L = L_cls + λL_reg
where L_cls is the classification loss (cross-entropy) and L_reg is the bounding box regression loss (smooth L1 loss).
Dataset and Preprocessing
The RSNA Pneumonia Detection Challenge dataset contains over 26,000 chest X-ray images from the National Institutes of Health Clinical Center. Each image includes:
- Patient ID and binary pneumonia classification target
- Bounding box annotations (x-min, y-min, width, height) for pneumonia regions
- Multiple bounding boxes per image when applicable
Data Preprocessing Pipeline
- Custom PyTorch dataset class for DICOM image handling
- Image resizing and tensor conversion
- Normalization and data augmentation
- Invalid bounding box filtering for data quality
- Train/validation split with minibatch loading
Results and Performance Comparison
| Method | Training Accuracy | Validation Accuracy | Training Loss | Validation Loss | Epochs |
|---|---|---|---|---|---|
| Method 1: Torch + Adam | 82.45% | 72.61% | 0.52 | 0.55 | 2 |
| Method 2: ImageNet + ResNet-50 | 77.34% | 76.81% | 0.5450 | 0.5151 | 1 |
| Method 3: PyTorch + SGD | 83.33% | 73.95% | 0.30 | N/A | 2 |
Method 3 loss convergence: Training loss decreased to approximately 0.3 over batch iterations
Key Findings
Best Overall Performance: Method 2 (Faster R-CNN with ImageNet pretrained ResNet-50) achieved the highest validation accuracy of 76.81%, demonstrating superior generalization despite training for only 1 epoch.
Transfer Learning Impact: ImageNet pretraining provided better initialization than COCO pretraining for this medical imaging task, likely due to the larger and more diverse ImageNet dataset.
Training Efficiency: Method 2 reached optimal performance in fewer epochs, suggesting that ImageNet features transfer effectively to medical image analysis tasks.
Computational Constraints: Limited computational resources restricted extensive hyperparameter exploration and prevented evaluation of additional architectures like YOLO.
Technical Achievements
- Multi-architecture Evaluation: Successfully implemented and compared three distinct Faster R-CNN variants with different backbones and optimization strategies
- Medical Image Pipeline: Developed robust preprocessing pipeline for DICOM chest X-ray images with bounding box handling
- Transfer Learning Analysis: Demonstrated effectiveness of ImageNet pretraining over COCO pretraining for medical imaging applications
- Performance Optimization: Achieved competitive accuracy (76.81% validation) on challenging medical dataset with limited computational resources
- Reproducible Framework: Created modular, extensible codebase for pneumonia detection that can be adapted for other medical imaging tasks
Clinical Relevance
This work demonstrates the feasibility of automated pneumonia detection systems that could assist radiologists in clinical settings. The achieved accuracy levels suggest potential for:
- Prioritizing urgent cases requiring immediate radiologist review
- Reducing diagnostic burden in resource-limited healthcare settings
- Providing consistent preliminary screening for large volumes of chest X-rays
- Supporting radiologist decision-making with automated opacity localization