Chest X-Ray Classification Using Deep Learning

Clone & Set Up

git clone https://github.com/your-username/Deep-Learning-Project.git
cd Deep-Learning-Project
pip install -r requirements.txt

---

Try Our Model

Test our final model on Hugging Face Spaces: DannyNet Demo

Running the Code

To train each model, navigate to the scripts directory and run the corresponding script. For example:

python scripts/dannynet.py

Evaluation results such as AUC and F1 scores will be printed in the console and logged to Weights & Biases (WandB) if your account is configured. The best model checkpoint will be saved in a models/<run_id> folder.

---

Project Overview

This project replicates and extends the findings from the paper: CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning.

We evaluate whether deep learning models, particularly CNNs and Transformers, can classify 14 chest pathologies from the NIH Chest X-ray dataset and potentially match or surpass expert radiologist performance.

---

Project Structure

Deep-Learning-Project/ 
  ├── eda.ipynb
  ├── scripts/ 
    ├── dannynet.py 
    ├── replicate_chexnet.py
    ├── vit_transformer.py 
  ├── XRay_app/ 
    ├── app.py
    ├── utils/
      ├── model_utils.py
      ├── preprocessing.py
  ├── test-images/
    ├── cardiomegaly.png
    ├── hernia-infiltration.png 
    ├── mass-nodule.png 
  ├── .gitignore 
  ├── requirements.txt 
  ├── README.md

---

Dataset

• Source: NIH Chest X-ray Dataset (Kaggle)
• Official Release: NIH Press Release
• Size: 112,120 X-ray images from 30,805 unique patients
• Labels: Multi-label classification of 14 thoracic diseases

Preprocessing

• Resizing to 224×224
• Normalization using ImageNet mean/std
• Data augmentation (e.g., rotation, brightness, contrast)
• Libraries used: torchvision.transforms
• Dataset filtering: Only includes images present in local folders and valid view positions (PA, AP)

---

Models

We present three models that were trained as separate python scripts located in the scripts folder

1. replicate_chexnet.py – Baseline Reimplementation of CheXNet

Faithful reimplementation of the original CheXNet architecture using standard PyTorch tools.

Key Features:

• Architecture: DenseNet-121 (pretrained)
• Classifier: Fully connected output for 14 diseases
• Loss: BCEWithLogitsLoss (standard binary cross-entropy)
• Optimizer: Adam with weight_decay=1e-5
• Scheduler: ReduceLROnPlateau (patience=1)
• Augmentations:
  • Resize(224)
  • RandomHorizontalFlip
• Evaluation: AUC-ROC and F1 (threshold=0.5)
• Filtering: Uses only PA/AP views as per the original paper

Training Strategy:

• Patient-level splitting
• Early stopping (patience=5)
• Epochs: 20 max

Hyperparameters:

• Batch size: 16
• Learning rate: 0.001
• Seed: 42

Purpose: Serves as a baseline for evaluating improvements from custom architectures like dannynet.py.

---

2. dannynet.py – Final Custom CNN Model (Best Performer)

A DenseNet-121-based CNN enhanced with Focal Loss and advanced augmentations to handle the class imbalance in the dataset.

Key Features:

• Architecture: DenseNet-121 (pretrained)
• Classifier: Fully connected layer outputting 14 logits
• Loss: Focal Loss (alpha=1, gamma=2)
• Optimizer: AdamW with weight_decay=1e-5
• Scheduler: ReduceLROnPlateau (patience=1, factor=0.1)
• Augmentations:
  • RandomResizedCrop(224)
  • RandomHorizontalFlip
  • ColorJitter
• Evaluation: AUC-ROC, F1 scores, optimal F1 thresholds per class

Training Strategy:

• Patient-level split to avoid data leakage
• Early stopping (patience=5)
• Epochs: 25 max

Hyperparameters:

• Batch size: 8
• Image size: 224x224
• Learning rate: 0.00005
• Seed: 42

Performance: Highest average AUC and F1 across all models; used in final deployment/demo.

---

3. vit_transformer.py – Vision Transformer for Chest X-Ray Classification

Explores transformers for medical image classification using Hugging Face's ViT-Base Patch16 224 model.

Key Features:

• Architecture: Vision Transformer (ViT) pretrained on ImageNet
• Loss: BCEWithLogitsLoss
• Optimizer: Adam (lr=0.0001, weight_decay=1e-5)
• Scheduler: ReduceLROnPlateau (patience=3)
• Preprocessing:
  • Uses ViTFeatureExtractor
  • Resize + normalize to transformer expectations

Evaluation:

• AUC-ROC and F1 (per class and average)
• Threshold = 0.5

Training Strategy:

• Patient-level split (80/10/10)
• 20 epochs max
• Early stopping (patience=5)

Batch Size: 16
Image Size: 224x224

Why Transformers? ViTs treat images as sequences of patches and apply self-attention to model global image features, which can be advantageous for complex medical images.

---

Performance vs older models and publications on Test AUC scores per disease

Pathology	original CheXNet	dannynet.py	vit_transformer.py	replicate_chexnet.py
Atelectasis	0.8094	0.817	0.774	0.762
Cardiomegaly	0.9248	0.932	0.89	0.922
Consolidation	0.7901	0.783	0.789	0.746
Edema	0.8878	0.896	0.876	0.864
Effusion	0.8638	0.905	0.857	0.883
Emphysema	0.9371	0.963	0.828	0.85
Fibrosis	0.8047	0.814	0.772	0.766
Hernia	0.9164	0.997	0.872	0.925
Infiltration	0.7345	0.708	0.7	0.673
Mass	0.8676	0.919	0.783	0.824
Nodule	0.7802	0.789	0.673	0.646
Pleural Thickening	0.8062	0.801	0.766	0.756
Pneumonia	0.768	0.74	0.713	0.656
Pneumothorax	0.8887	0.875	0.821	0.827

---

Average metrics across all diseases for each model

Metric	DannyNet	ViT Transformer	Replicate CheXNet
Loss	0.0416	0.1589	0.1661
AUC	0.8527	0.7940	0.7928
F1	0.3861	0.1114	0.0763

---

📊 F1 Score Comparison for Each Model

Disease	DannyNet	ViT Transformer	Replicate CheXNet
AVERAGE	0.386	0.111	0.076
Atelectasis	0.421	0.127	0.026
Cardiomegaly	0.532	0.264	0.423
Consolidation	0.226	0	0
Edema	0.286	0.004	0
Effusion	0.623	0.427	0.459
Emphysema	0.516	0.079	0
Fibrosis	0.127	0	0
Hernia	0.750	0	0
Infiltration	0.395	0.193	0.061
Mass	0.477	0.213	0.079
Nodule	0.352	0.041	0
Pleural Thickening	0.258	0	0
Pneumonia	0.082	0	0
Pneumothorax	0.360	0.211	0.021

---

Test-Images

Folder that contains labeled chest X-ray PNG files for the user to easily download and test on the Hugging Face Streamlit app.

---

project_EDA.ipynb

Jupyter Notebook that conducts Exploratory Data Analysis such as how many different diseases are present in the dataset and the proportion of each disease in the dataset.

Methodology

• Data Preprocessing: Resize, normalize, and augment X-ray images
• Model Selection: CNN and Transformer variants
• Training & Validation: Patient-level splits, loss monitoring
• Evaluation: Per-class AUC-ROC and F1 metrics
• Comparison: Benchmarks against original CheXNet results