EffiDerm: An Efficient Deep Learning Model for Skin Cancer Prediction
📌 Overview
EffiDerm is a lightweight Convolutional Neural Network (CNN) designed for skin cancer detection and classification. It classifies 7 types of skin lesions from dermoscopic images with high accuracy while maintaining a low computational footprint.
Unlike heavy models such as ResNet-50 or VGG16, EffiDerm has only ~1.4M parameters, making it suitable for mobile deployment and real-time usage in resource-limited healthcare environments.
✨ Features
🔹 Lightweight CNN with only ~1.4M parameters
🔹 93.07% classification accuracy
🔹 Inference time ~47 ms per image
🔹 Memory footprint ~5.6 MB (mobile-friendly)
🔹 Class balancing with SMOTE for rare lesion types
🔹 Data augmentation to improve generalization
🔹 Grad-CAM visualizations for explainability
🔹 Deployable in VS Code, local GPUs, or mobile frameworks
📊 Dataset
We use the HAM10000 dataset (Human Against Machine with 10,000 training images).
Total Images: 10,015 dermatoscopic images
Classes: 7 skin lesion categories
Format: RGB, 600×450 px (resized to 64×64 px)
Source: Harvard HAM10000 Dataset
🛠️ Tech Stack
Programming Language: Python 3.8
Frameworks: TensorFlow 2.x, Keras
Libraries: NumPy, Pandas, Matplotlib, Seaborn, Scikit-learn, Imbalanced-learn (SMOTE), OpenCV
Development Environment: Visual Studio Code
⚙️ Methodology
Preprocessing
Resize all images to 64×64×3
Normalize pixel values to [0, 1]
One-hot encode labels
Apply SMOTE to handle class imbalance
Data Augmentation
Random rotation, zoom, shift
Horizontal & vertical flipping
Brightness adjustment
Model Training
CNN with Conv2D, MaxPooling, Dropout, BatchNormalization
Optimizer: Adamax
Loss: Categorical Crossentropy with Focal Loss
Epochs: 50, Batch size: 32
Evaluation
Metrics: Accuracy, Precision, Recall, F1-score
Confusion Matrix
Training/Validation Curves
Grad-CAM Heatmaps
📈 Results
Accuracy: 93.07%
Inference Time: ~47 ms
Parameters: ~1.4M
Memory: ~5.6 MB
🌍 Applications
Clinical decision-support tool for dermatologists
Mobile-based early skin cancer screening app
Telemedicine integration for rural healthcare
Educational tool for dermatology students
Dataset contains only dermoscopic images (not smartphone-quality photos)
Model performance depends on dataset quality and diversity
Requires further validation in real-world clinical settings
🔮 Future Scope
Integration with explainable AI (better Grad-CAM visualizations)
Deployment in mobile and edge devices
Expansion to include real-world smartphone datasets
Cloud-based teledermatology platforms
👨💻 Author
Agil Kannan
🎓 B.Tech IT Student (2021–2025)
✨ If you like this project, don’t forget to ⭐ the repo!