---
title: SkinAI Diagnostics
emoji: 🔬
colorFrom: blue
colorTo: purple
sdk: docker
app_port: 7860
pinned: false
---

# Skin Cancer Classifier API

SkinAI Diagnostics — a FastAPI-powered web application for skin cancer image classification using deep learning.

---

## Overview

This project delivers a professional web interface for classifying dermatoscopic images. It enables users to upload a skin lesion image and obtain a prediction with calibrated confidence scores, alongside links to technical details of the underlying model.

---

## Repository Structure

```
skin-cancer-api/
├── app/
│   ├── main.py                # FastAPI application entry point
│   ├── predict.py             # Image preprocessing and prediction logic
│   ├── model_loader.py        # Loads the trained EfficientNetV2S model
│   ├── model/                 # Local model storage (auto-downloaded if missing)
│   │   └── efficientnetv2s.h5 # Pretrained model file
├── test_images/               # Sample images for testing the app
│    └── ...
├── requirements.txt           # Python dependencies
├── Dockerfile                 # Docker setup for deployment
├── .dockerignore              # Docker ignore rules
└── README.md                  # Project documentation (this file)
```

---

## Features

- Seven-class classification of dermatoscopic images:
  - Actinic Keratoses and Intraepithelial Carcinoma (AKIEC)
  - Basal Cell Carcinoma (BCC)
  - Benign Keratosis-like Lesions (BKL)
  - Dermatofibroma (DF)
  - Melanoma (MEL)
  - Melanocytic Nevi (NV)
  - Vascular Lesions (VASC)
- Professional web interface built with FastAPI and Jinja2
- Temperature Scaling (T-scaling) for calibrated probabilities
- Technical transparency: model architecture, training setup, and metrics
- Confidence visualization with a probability chart
- Sample images available in `test_images/`

---

## Model Details

- Base Model: EfficientNetV2S (20.5 million parameters)
- Dataset: [HAM10000](https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/DBW86T)
- Calibration: Temperature Scaling (optimal T=2.77)
- Performance:
  - Accuracy: 0.88
  - Macro F1-score: 0.80
  - Expected Calibration Error (ECE): 0.022 (after T-scaling)

For full technical details, see the Model Technical Information section in the app.

---

## Model Download from Hugging Face

The model file `efficientnetv2s.h5` is hosted on [Hugging Face Hub](https://huggingface.co/Miguel764/efficientnetv2s-skin-cancer-classifier) and is automatically downloaded the first time the app runs.

How it works:

- Expected local path: `app/model/efficientnetv2s.h5`
- On startup, `app/model_loader.py` checks for the file
- If missing, it is downloaded via `huggingface_hub` and saved to `app/model/`
- Subsequent runs load the local copy

Note for Docker users: When running inside a container, the downloaded model is stored inside the container’s filesystem. Mount a volume if you need to persist it on the host.

Manual download link (optional): https://huggingface.co/Miguel764/efficientnetv2s-skin-cancer-classifier

---

## Installation & Usage

### 1) Clone the repository

```sh
git clone https://github.com/yourusername/skin-cancer-api.git
cd skin-cancer-api
```

### 2) Install dependencies

```sh
pip install -r requirements.txt
```

Note: Requires Python 3.8+ and TensorFlow 2.10.0 (GPU recommended).

### 3) Run the application

```sh
uvicorn app.main:app --reload
```

The app will be available at http://localhost:8000

Access the API docs at http://127.0.0.1:8000/docs

### 4) Try with sample images

Use the images in the `test_images/` folder to test the classifier.

---

## Docker Deployment

Build and run the app in a container:

```sh
docker build -t skin-cancer-api .
docker run -p 8000:8000 skin-cancer-api
```

---