🧬 Cervical Cancer Classification using NAS (DARTS) vs ResNet

📌 Overview

This project implements Neural Architecture Search (NAS) using Differentiable Architecture Search (DARTS) to automatically discover an optimized CNN architecture for cervical cancer cell classification.
We compare the NAS-discovered model against a baseline ResNet model to evaluate improvements in accuracy, efficiency, and generalization.

The dataset used is SIPaKMeD, which contains five classes of cervical cell images.

---

🎯 Project Goals

• Automate CNN architecture design using DARTS.
• Compare NAS-generated architecture vs ResNet baseline.
• Evaluate on key metrics: Accuracy, F1-score, Model Size, Inference Time.
• Visualize performance using Confusion Matrix and ROC curves.

---

📂 Dataset Structure

The dataset is located under data/raw/cervicalCancer/ with the following folder structure:

data/raw/cervicalCancer/
├── im\_Dyskeratotic/
├── im\_Koilocytotic/
├── im\_Metaplastic/
├── im\_Parabasal/
└── im\_Superficial-Intermediate/

---

🏗 Project Structure

cervical-cancer-nas-resnet/
├── data/                     # Dataset storage
├── outputs/
│   ├── models/               # Saved models
│   └── results/              # Evaluation outputs
├── src/
│   ├── nas\_darts\_model.py    # DARTS model architecture
│   ├── nas\_darts\_search\_space.py
│   ├── nas\_genotypes.py      # Genotype storage
│   ├── nas\_train\_darts.py    # DARTS training
│   ├── train\_resnet.py       # ResNet baseline training
│   ├── evaluate\_darts.py     # Evaluation for DARTS
│   ├── evaluate\_resnet.py    # Evaluation for ResNet
│   └── data\_utils.py         # Data loading utilities
├── requirements.txt
└── README.md

---

⚙️ Installation

# Clone the repository
git clone https://github.com/villwin007/cervical-cancer-nas-resnet.git
cd cervical-cancer-nas-resnet

# Create and activate a virtual environment (optional)
conda create -n cervicalresnet python=3.10 -y
conda activate cervicalresnet

# Install dependencies
pip install -r requirements.txt

---

🚀 Usage

1️⃣ Train DARTS Model

python src/nas_train_darts.py

This will:

• Train the DARTS-based architecture search model.
• Save the best discovered model in outputs/models/darts_final.pth.

2️⃣ Train ResNet Baseline

python src/train_resnet.py

This will:

• Train a standard ResNet model.
• Save the model in outputs/models/resnet_final.pth.

3️⃣ Evaluate DARTS

python src/evaluate_darts.py

Generates:

• Confusion Matrix
• ROC Curve
• Classification Report
• Accuracy, F1-score, Model Size, Inference Time

4️⃣ Evaluate ResNet

python src/evaluate_resnet.py

Generates:

• Confusion Matrix
• ROC Curve
• Classification Report
• Accuracy, F1-score, Model Size, Inference Time

---

📊 Results

Metric	ResNet Baseline	NAS (DARTS)
Accuracy	78.43%	81.06%
F1-score (macro)	0.77	0.81
Model Size (MB)	44.3	18.7
Inference Time	7.9 ms/image	5.2 ms/image

---

📈 Visualizations

Confusion Matrix (DARTS)

ROC Curve (DARTS)

Confusion Matrix (Resnet)

ROC Curve (Resnet)

---

🧪 Key Findings

• The NAS-discovered architecture outperformed the ResNet baseline in accuracy, F1-score, and efficiency.
• Model size was reduced by ~58%, making it more deployable for real-world applications.
• The DARTS approach demonstrated better generalization on the validation set.

---

📜 Citation

If you use this work, please cite:

@misc{cervical-nas-resnet,
  author = {Dhanush Saravanan},
  title = {Cervical Cancer Classification using NAS (DARTS) vs ResNet},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/<villwin007>/cervical-cancer-nas-resnet}
}

---

📧 Contact

For questions, feel free to reach out:

• Author: Dhanush Saravanan
• Email: villwin11@gmail.com
• GitHub: villwin007

---