Comparative Analysis of CNNs vs. Vision Transformers on Chest X-Rays

📄 Abstract

This project conducts a comprehensive comparative analysis of Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) for the classification of thoracic diseases. Using the VinDr-CXR (Adult) and VinDr-PCXR (Pediatric) datasets, we evaluated state-of-the-art architectures including ResNet, EfficientNet, ConvNeXt, Swin Transformer, and CoAtNet.

Our findings indicate that while CNNs (like CoAtNet) excel in capturing fine-grained local features, Vision Transformers (like Swin) offer superior performance in modeling long-range dependencies, achieving competitive AUC scores across both multi-label and multi-class tasks.

---

🚀 Key Results

We implemented a robust training pipeline including CLAHE preprocessing for data augmentation.

1. Adult Dataset (VinDr-CXR) - Multi-Label

Model	Accuracy	AUC	Precision
Swin Transformer	95.22%	0.9328	0.5959
CoAtNet	95.09%	0.9302	0.7021
ConvNeXt	95.43%	0.9267	0.6999

2. Pediatric Dataset (VinDr-PCXR)

• Multi-Label: CoAtNet achieved the highest AUC of 0.7316, slightly outperforming Swin Transformer (0.7259).
• Multi-Class: PVT and EfficientNet showed the highest accuracy (~67%), highlighting the difficulty of pediatric classification compared to adult X-rays.

Full results table available in assets/tables/Performance_Metrics.csv.

---

📂 Repository Structure

The repository contains representative notebooks for the top-performing models discussed in the thesis.

CXR-Classification-ViT-vs-CNN/
│
├── assets/
│   ├── figures/
│   │   ├── architectures/   # Block diagrams of Swin, DaViT, CoAtNet, etc.
│   │   ├── statistics/      # Class distribution charts
│   │   └── preprocessing/   # Pipeline visualization
│   ├── samples/             # Sample X-rays from the datasets
│   └── tables/              # Raw result CSV
│
├── docs/                    # Full Thesis PDF
│
├── notebooks/
│   ├── VinDr_CXR/           # Adult Dataset (Unbalanced Multi-label)
│   ├── VinDr_PCXR/          # Pediatric Dataset (Balanced & Unbalanced)
|   ├──Model_Architectures.ipynb    # Contains all model definitions
|   ├──Ensemble_CoAtNet_Swin.ipynb  # The Ensemble (of CoAtNet and Swin) Inference file
|
└── README.md

📊 Dataset Overview

To address the significant class imbalance in medical imaging, we analyzed the label distributions and curated specific subsets for training. The tables below detail the sample counts for the Unbalanced (Original) and Balanced (Selected) sets used in our experiments.

Adult Dataset (VinDr-CXR)	Pediatric Dataset (VinDr-PCXR)
Class distribution for Multi-label classification	Comparison of Original vs. Balanced subsets

---

🧠 Model Architectures

We analyzed the internal mechanisms of both CNNs and ViTs. Click below to view the architectural details.

🔍 View Swin Transformer & DaViT (Vision Transformers)

Swin Transformer (Shifted Windows)	DaViT (Dual-Attention Mechanism)

🔍 View CoAtNet & ConvNeXt (Hybrid & Modern CNN)

CoAtNet (Hybrid)	ConvNeXt (Pure CNN)

---

🖼️ Data Preprocessing Pipeline

We implemented a 3-stage preprocessing pipeline to enhance feature visibility in low-contrast X-rays:

Step 1: Original	Step 2: VOI LUT Enhancement	Step 3: Min-Max Norm and CLAHE

---

🛠️ Usage

To reproduce the experiments:

1. Clone the Repository:

   git clone https://github.com/CellerCity/CXR-Classification-ViT-vs-CNN.git

2. Install Dependencies:

   pip install -r requirements.txt

3. Dataset Setup (Crucial):

   • Download: Obtain the VinDr-CXR and VinDr-PCXR datasets.
     • Note: VinDr-PCXR requires signing a data use agreement for access.
   • Update Paths: The notebooks currently use Kaggle directory paths. You must update the train_data and/or test_data variables in the first cell of the notebooks to point to your local dataset location.

4. Run Experiment: Open any notebook in notebooks/ (e.g., VinDr_CXR/Swin_CXR_Unbalanced_Multilabel.ipynb). The training loop is self-contained.

   • Note: For ensemble inference (combining CoAtNet + Swin), refer to notebooks/Ensemble_CoAtNet_Swin.ipynb.

---

👥 Contributors

• Aman Nasim
• Bobbili Sunny Rishy Vardhan
• Kalakonda Prudhvi Raj

Supervised by Dr. Srilatha Chebrolu, NIT Andhra Pradesh (2024).