|
|
--- |
|
|
title: DeepDerma |
|
|
emoji: 🧴 |
|
|
colorFrom: blue |
|
|
colorTo: pink |
|
|
sdk: gradio |
|
|
sdk_version: 5.38.0 |
|
|
app_file: app.py |
|
|
pinned: true |
|
|
short_description: Detect skin cancer early with powerful AI |
|
|
--- |
|
|
|
|
|
# 🩺 DeepDerma: Skin Lesion Classification App |
|
|
|
|
|
Welcome to **DeepDerma**, a simple yet powerful AI tool that helps identify **7 common skin lesions (abnormal injury or disease)** from clinical dermatoscopic images. Upload a skin image, and DeepDerma will predict the most likely diagnosis — assisting in early detection and educational awareness. |
|
|
|
|
|
|
|
|
|
|
|
--- |
|
|
|
|
|
## 🔍 How It Works |
|
|
|
|
|
Just upload a skin lesion image, and our AI model will: |
|
|
- Preprocess the image |
|
|
- Classify it into one of 7 dermatological categories |
|
|
- Return the top predicted class with confidence scores |
|
|
|
|
|
The model is built using **EfficientNet-B2** and trained on the **DermMNIST** dataset from MedMNIST. |
|
|
|
|
|
--- |
|
|
|
|
|
## 🧪 Performance Summary |
|
|
|
|
|
| Metric | Value | |
|
|
|--------------|-----------| |
|
|
| Test Accuracy | 73.3% | |
|
|
| AUC Score | 0.91 | |
|
|
| Top Class F1 | 0.86 (Nevus - NV) | |
|
|
| Minority Class F1 | 0.53 | |
|
|
|
|
|
## Competitiveness |
|
|
|
|
|
> our results outperforms benchmarks such as ResNet-18, ResNet-50 in terms of accuracy and is competitive in AUC scores |
|
|
|
|
|
Despite class imbalance, the model performs well on high-priority categories like melanoma (MEL) and nevi (NV) thanks to AUC-based training. |
|
|
|
|
|
--- |
|
|
|
|
|
## 🧠 Model Details |
|
|
|
|
|
- **Architecture**: [EfficientNet-B2](https://arxiv.org/abs/1905.11946) |
|
|
- **Fine-tuned** on: DermMNIST (medmnist v2) |
|
|
- **Input size**: 224 × 224 |
|
|
- **Optimizer**: Adam, LR = 1e-4 |
|
|
- **Scheduler**: ReduceLROnPlateau |
|
|
- **Augmentations**: Random flip, rotation, color jitter |
|
|
- **Class balancing**: Weighted loss + WeightedRandomSampler |
|
|
- **Metric used**: AUC (Area Under ROC Curve) for better performance on imbalanced classes |
|
|
|
|
|
--- |
|
|
|
|
|
## 📊 Dataset: DermMNIST |
|
|
|
|
|
- **Source**: [MedMNIST v2](https://medmnist.com/) |
|
|
- **Images**: 10,015 dermatoscopic RGB images (28×28, resized to 224×224) |
|
|
- **Classes**: 7 types of skin lesions |
|
|
- **Split**: |
|
|
- Train: 7,007 images |
|
|
- Val: 1,003 images |
|
|
- Test: 2,005 images |
|
|
|
|
|
--- |
|
|
|
|
|
## 🧬 Target Classes (With Description) |
|
|
|
|
|
| Label | Name (Short) | Description | |
|
|
|-------|--------------|-------------| |
|
|
| 0 | **AKIEC** | Actinic keratoses / Intraepithelial carcinoma – pre-cancerous skin lesions | |
|
|
| 1 | **BCC** | Basal Cell Carcinoma – common and locally invasive skin cancer | |
|
|
| 2 | **BKL** | Benign Keratosis-like lesions – non-cancerous growths (seborrheic, solar, etc.) | |
|
|
| 3 | **DF** | Dermatofibroma – benign skin nodules caused by overgrowth of fibrous tissue | |
|
|
| 4 | **MEL** | Melanoma – the most dangerous type of skin cancer; early detection critical | |
|
|
| 5 | **NV** | Melanocytic Nevi – common moles, typically benign | |
|
|
| 6 | **VASC** | Vascular Lesions – angiomas, hemorrhages, and similar blood vessel-related growths | |
|
|
|
|
|
--- |
|
|
|
|
|
## 🚀 How to Run |
|
|
|
|
|
This Space runs using **Gradio**. No setup needed — just: |
|
|
|
|
|
1. Click the upload button |
|
|
2. Select or drag an image |
|
|
3. View the predicted class and probabilities |
|
|
|
|
|
--- |
|
|
|
|
|
## 🧾 Files Included |
|
|
|
|
|
- `app.py` — Gradio interface |
|
|
- `model.py` — Model architecture and prediction pipeline |
|
|
- `requirements.txt` — Dependencies |
|
|
- `fine_tuned_effnetb2_dermamnist.pth` — Trained model weights |
|
|
|
|
|
--- |
