CLARITY - Models Repository

CLARITY: A Multimodal Deep Learning Framework for Chest X-ray Diagnosis and Automated Radiology Report Generation

📋 Overview

This repository contains the machine learning model development, training, and XAI implementation for the CLARITY framework. It houses all model architectures, training pipelines, explainable AI implementations, and evaluation code for the automated chest X-ray diagnosis system.

🎯 Project Goal

Develop a clinically-validated deep learning framework to assist radiologists in chest X-ray diagnosis by:

• Multi-label classification of 14 chest pathologies
• Providing explainable predictions through 5 XAI methods
• Generating structured clinical reports
• Addressing radiologist shortage in resource-constrained settings

---

---

🚀 Quick Start

Installation

# Clone repository
git clone https://github.com/harshak103/CLARITY.git
cd CLARITY

# Create virtual environment
conda create --prefix ./CLARITY.env python==3.10 -y
conda activate ./CLARITY.env

# Install dependencies
pip install -r requirements.txt

Dataset Setup

# Download NIH ChestX-ray14 dataset
# From: https://www.kaggle.com/datasets/nih-chest-xrays/data

# Expected structure:
# data/raw/
#   ├── images_001/
#   ├── images_002/
#   ├── .....
#   ├── images_012/
#   ├── Data_Entry_2017.csv (Metadata file)
#   └── train_val_list.txt  (Split configuration)

📊 Model Performance

DenseNet121 (Selected Production Model)

• Accuracy: 93.4%
• AUC: 0.9154
• Parameters: 7M
• Training Time: ~40 min/epoch (20 epochs for baseline)
• Inference Time: 150ms per image (CPU), 45ms (GPU)

ResNet152 (Comparison Baseline)

• Accuracy: 93.36%
• AUC: 0.8337
• Parameters: 60M
• Inference Time: 200ms per image

Per-Pathology Performance

Pathology	Precision	Recall	F1-Score	AUC
Pneumonia	0.78	0.54	0.42	0.82
Effusion	0.81	0.56	0.43	0.80
Cardiomegaly	0.72	0.48	0.38	0.75
Infiltration	0.65	0.35	0.28	0.68
Atelectasis	0.58	0.31	0.18	0.62

---

🧠 XAI Implementations

Five Attribution Methods

1. Grad-CAM++ - Second-order gradients for multifocal pathologies
2. LayerCAM - Preserves intermediate layer information
3. Score-CAM - Gradient-independent, eliminates artifacts
4. Saliency Maps - Pixel-level gradient sensitivity
5. Integrated Gradients - Axiomatic path-based attribution

📈 Training Configuration

Optimal Hyperparameters

learning_rate: 0.0001
batch_size: 32
epochs: 50
optimizer: Adam
loss_function: Weighted Focal Loss
weight_decay: 1e-4
dropout: 0.3

Data Augmentation

augmentation_strategy: Domain-Specific
- horizontal_flip: 0.5
- rotation: ±15°
- color_jitter: 0.2
- brightness: ±10%

Loss Function

# Weighted Focal Loss for class imbalance
# Improved rare pathology detection by 8-15%

---

🔬 Experimental Results

Model Variations

• Standard DenseNet121: 93.4% accuracy
• Deep DenseNet169: 92.81% accuracy (2.3× parameters)
• Lightweight DenseNet: 92.15% accuracy (40% fewer parameters)

Augmentation Impact

• Aggressive: +0.015 AUC (35% training overhead)
• Domain-specific: +0.08-0.12 on low-quality images
• Conservative: Minimal improvement

Loss Function Comparison

• Binary CE: 0.82 AUC
• Weighted BCE: 0.84 AUC
• Focal Loss: 0.85 AUC
• Weighted Focal Loss: 0.87 AUC ✓ Selected

---

📚 Dataset Information

NIH ChestX-ray14 Dataset

• Size: 112,120 frontal-view X-ray images
• Pathologies: 14 disease labels
• Class Distribution: Highly imbalanced (some pathologies <1%)
• Resolution: Variable (typically 1024×1024)
• Data Split: 70% train, 20% validation, 10% test

Pathologies Included

1. Atelectasis
2. Cardiomegaly
3. Effusion
4. Infiltration
5. Mass
6. Nodule
7. Pneumonia
8. Pneumothorax
9. Consolidation
10. Edema
11. Emphysema
12. Fibrosis
13. Pleural Thickening
14. Hernia

---

🔍 Key Findings

Class Imbalance Handling

• Inverse frequency weighting effective for rare pathologies
• Weighted focal loss improved F1 by 8-15% for rare classes
• Learned parameters enable automatic per-batch adjustment

Model Selection Trade-offs

• DenseNet121: 8.6× fewer parameters, marginally better accuracy
• ResNet152: Higher capacity, prohibitive for deployment
• Production choice: DenseNet121 (efficiency + performance)

Generalization

• Domain-specific augmentation: +8-12% on degraded images
• Strong performance on high-quality radiographs
• Consistent across image quality levels

---

🛠️ Requirements

Core Dependencies

• PyTorch 2.0+
• CUDA 11.8 (GPU support)
• NumPy, Pandas, Scikit-learn
• OpenCV, Pillow (Image processing)
• Matplotlib, Plotly (Visualization)

Full Requirements

See requirements.txt for complete dependency list

torch==2.0.0
torchvision==0.15.0
numpy==1.24.3
pandas==2.0.2
scikit-learn==1.2.2
opencv-python==4.7.0
matplotlib==3.7.1
plotly==5.14.0

---

📝 Training Logs

All training runs are logged with:

• Loss curves (training & validation)
• Per-pathology metrics
• Hyperparameter configurations
• Computational resource usage
• XAI computation times

Access logs in logs/ directory

---

🤝 Contributing

Contributions welcome! Areas for improvement:

• 3D volumetric analysis
• Multi-modal integration (CT, MRI)
• Real-time inference optimization
• Additional XAI methods (LIME, SHAP)

---

📞 Contact & Support

• Author: Harsha K
• GitHub: @harshak103
• Issues: Report via GitHub Issues tab
• Email: [email protected]

---

MIT License - See LICENSE file for details

---

⭐ Acknowledgments

• NIH for ChestX-ray14 dataset
• PyTorch team for deep learning framework
• Research team for guidance and feedback

---

Last Updated: November 30, 2025 Status: ✅ Production-Ready