Computer-Aided Skin Cancer Diagnosis (ResNet + Web App)

---

📌 About the Project

I'm Nishita. This project began as a part of our Deep Learning elective coursework in my 4th semester (Jan–May 2025). I’ve always been fascinated by the intersection of AI and healthcare, so I took up automated skin cancer classification a topic I quickly realized was much more complex than it looked on the surface.

I initially started with standard CNN models, but the results were unstable, especially for rare classes like vascular lesions. After a lot of experimentation, I shifted to ResNet-50 a deeper model with residual blocks that made training much more stable and gave significantly better results, especially under image noise or poor contrast.

This was more than just a model training exercise from data cleaning, preprocessing, class imbalance correction, to late-night debugging and waiting for all 50 epochs to finally run — it was a real deep dive into applied deep learning.

To complete the pipeline, I wrapped the model in a Flask web app, designed a simple frontend, and deployed it — so anyone could upload an image and get predictions along with Grad-CAM heatmaps to interpret the results.

---

🎯 What I Wanted to Achieve

• Train a CNN to classify 7 different types of skin lesions
• Improve performance on underrepresented classes (e.g. AKIEC, DF)
• Implement Grad-CAM for visual interpretability
• Deploy a working diagnostic demo via Flask

---

✅ What I Actually Achieved

• 🧠 Trained and fine-tuned ResNet-50 with class-weighted loss and preprocessing
• ⚙️ Achieved 84.3% accuracy and 91% sensitivity on test set
• 📈 Compared with MobileNetV2 and ensemble versions
• 🔬 Used Grad-CAM to explain model predictions
• 🌐 Deployed a fully working Flask web app with image upload and live results
• 🧪 Experimented with dropout tuning, early stopping, and noise robustness tests

---

🔍 Technical Deep Dive

• Model Choice: Started with a vanilla CNN but it overfit quickly. ResNet-50 offered skip connections that helped avoid vanishing gradients and trained better on medical images.
• Preprocessing:
  • All images converted to grayscale
  • Histogram equalization to improve lesion visibility
  • Resized to 224x224
• Training Details:
  • Framework: TensorFlow + Keras
  • Optimizer: Adam, LR = 1e-4 with ReduceLROnPlateau
  • Loss: CategoricalCrossentropy with class reweighting
  • Regularization: Dropout (0.3), EarlyStopping
• Data Augmentation:
  • Horizontal/vertical flips
  • Random zoom & brightness
• Validation Strategy:
  • 80/10/10 split (train/val/test)
  • Used stratified sampling to preserve class ratios

---

⚙️ Tech Stack

Category	Tools Used
💻 Languages	Python, HTML, CSS, JavaScript
🔍 Deep Learning	TensorFlow, Keras, tf-keras-vis
🧰 Tools/Libraries	NumPy, Matplotlib, OpenCV, Flask
📊 Visualization	Grad-CAM for interpretability
🖥 Deployment	Flask + GitHub Pages (for static frontend)

---

📁 File Structure

📦skin-cancer-diagnosis
├── app/
│   ├── app.py              # Flask app main server
│   ├── model_loader.py     # Load model and Grad-CAM logic
│   ├── utils.py            # Image preprocessing
│   ├── templates/
│   │   └── index.html      # Frontend UI
│   └── static/
│       └── style.css       # CSS styles
├── models/
│   └── resnet50_model.h5   # Trained model weights
├── notebooks/
│   └── training_resnet.ipynb  # Full training notebook
├── requirements.txt
├── README.md
└── run.sh

---

🧪 Dataset Info

• Source: ISIC 2018 Challenge Dataset

• Classes (7 total):

  • Melanoma (MEL)
  • Nevus (NV)
  • Basal Cell Carcinoma (BCC)
  • Actinic Keratoses (AKIEC)
  • Benign Keratosis (BKL)
  • Dermatofibroma (DF)
  • Vascular lesions (VASC)

• Challenge: Very imbalanced dataset — classes like VASC and DF had <5% representation.

• Solution: Applied class-weighting in the loss function and oversampling in training batches.

---

📈 Results Summary

Model	Accuracy	Sensitivity	Recall (Rare)	Test Robustness
ResNet-50	84.3%	91.0%	83.2%	✅ High
MobileNetV2	75.1%	81.2%	88.4%	⚠️ Unstable
Ensemble (Mob)	79.5%	85.0%	91.0%	⚠️ Mixed

---

🧠 Grad-CAM Visualizations

• Implemented Grad-CAM to visualize which parts of the lesion image influenced the model’s prediction.
• Made it part of the live web app, not just the notebook!
• Used tf-keras-vis to extract final conv layer activations and overlaid on input image.

---

🌐 Try It Live

💡 Try the live frontend: 🔗 Skin Cancer Diagnosis Web UI

(Frontend only — model hosted locally for now. Full deployment coming soon!)

---

🚀 How to Run Locally

git clone https://github.com/yourusername/skin-cancer-diagnosis.git
cd skin-cancer-diagnosis
pip install -r requirements.txt
python app/app.py

Open http://localhost:5000 in your browser to use the diagnostic app.

---

🧰 Deployment Notes

To deploy:

1. Host app/ on Render or Railway (recommended)

2. Use run.sh as startup script

3. Static UI deployed via GitHub Pages:

   • templates/index.html and static/style.css in gh-pages branch

---

🔮 Future Work

• Convert into Streamlit / React-based frontend for faster loading
• Add patient database and login feature
• Dockerize the entire project for deployment on any platform
• Add clinical validation with real feedback from medical professionals

---

🙏 Acknowledgements

• Thanks to ISIC Archive for the open-access dataset
• TensorFlow and Keras documentation for making model tuning bearable
• tf-keras-vis for easy Grad-CAM implementation
• Everyone in our Deep Learning class who helped debug weird tensor shapes at 2am 🙃

---