An applied machine learning project focused on classifying skin cancer images into benign and malignant cases using deep convolutional neural networks. This project explores multiple CNN architectures, data augmentation strategies, optimisers, and hyperparameter tuning — with EfficientNetB0 achieving state-of-the-art results.
🎓 Built as part of the EEEM068 – Applied Machine Learning coursework at the University of Surrey (2024).
- 🔍 Dataset: Melanoma Skin Cancer Dataset (Kaggle)
- 📸 10,605 high-resolution dermoscopic images
- 🧪 Classes: Binary –
0(Benign) and1(Malignant) - ⚖️ Balanced dataset with no missing values
| Model | Accuracy (%) | F1 Score |
|---|---|---|
| ResNet34 | 92.46 | 0.9243 |
| ResNeXt101 | 92.55 | 0.9253 |
| MobileNetV2 | 90.94 | 0.9093 |
| EfficientNetB0 | 96.80 | 0.9679 |
EfficientNetB0 was selected as the final model for its optimal performance vs resource efficiency.
🏆 Our best-performing model, EfficientNetB0, achieved the following:
| Metric | Value |
|---|---|
| Accuracy | 96.80% |
| F1 Score | 0.9679 |
| Precision | 0.9751 |
| Recall | 0.9612 |
| AUC-ROC | 0.979 |
✅ Model trained using Adam optimizer, lr = 0.001, batch size = 64, for 25 epochs
✅ 3-fold validation was used to ensure generalizability
✅ Dataset: Melanoma Skin Cancer Dataset (Kaggle)
- Resized to 224x224
- Normalization (ImageNet mean/std)
- RandomHorizontalFlip, Rotation (±15°), ColorJitter, RandomResizedCrop
- Loss:
BCELoss - Optimisers tested:
Adam(best) andSGD - Grid search for:
- Learning rate (
[0.001, 0.01, 0.1]) - Batch sizes (
[16, 32, 64])
- Learning rate (
- Best combo:
lr=0.001,batch_size=64
| Training & Validation Loss | Confusion Matrix |
|---|---|
 |
 |
These maps show the model learning important lesion features like border irregularity and asymmetry—core visual traits used by dermatologists.
- 📚 Python, PyTorch, NumPy, Matplotlib
- 🧠 CNN Architectures: ResNet, MobileNet, EfficientNet
- 📁 Jupyter Notebook for training and evaluation
- Incorporate multimodal data (e.g. clinical metadata)
- Explore unsupervised melanoma detection
- Integrate explainable AI features for clinical transparency
- Taman Bachani
- Azizbek Fatkhullaev
- Aptha Sara Mohan
- Yasar Efe Pekgoz
For collaboration, reproduction, or research inquiry:
📧 workwith.taman@gmail.com
🔗 LinkedIn
MIT License