This is the source code for a deep learning based skin cancer detection android app. The model has been built using fastai deep learning library which is a high level api for pytorch. A flask API has also been implemented for cloud-based inference. The classifier has been trained and validated on Kaggle MNIST HAM10000 dataset which contains 10015 images of seven categories of pigmented skin lesions. As a preprocessing step, I have applied random undersampling to data to alleviate the class-imbalance problem. The classifier has been trained with transfer learning technique using a pretrained Densenet169 model. The final classifer achieved an accuracy of 91.2% and a F1-score of 91.7% on validation data. You can check out the jupyter notebook that goes along to follow all the steps which have been taken to build the model.
- Python 3.6
- Fastai 1.0.52
- Flask
- Gunicorn
- SquareCamera
- Volley
In order to setup flask API first run sudo pip install -r requirements.txt to install the required dependencies. Then launch the app by running python app.py. When you take a photo of skin lesion using the android app, a base64 encoding of the image will be sent to the API at http://localhost:8008. Once you set up the API, download Android Studio and then import the android app. Since I've placed the dependencies in the build.gradle file, they should be automatically downloaded.