Melanoma diagnosis service

Identify melanoma in lesion images

This repository holds the source code for my master thesis "Model as a Service: Development of a prototype for computer-aided skin cancer diagnosis".

To identify a melanoma, several neural networks were trained based on the Kaggle dataset from the "SIIM-ISIC Melanoma Classification" competition: https://www.kaggle.com/c/siim-isic-melanoma-classification/overview

The training process as well as the jupyter notebooks for all models can be found here: https://www.comet.ml/saschamet/master-thesis

A Kaggle notebook showing the training process of a single EfficientNet B5 model is available here: https://www.kaggle.com/saschamet/melanoma-efficientnetb5-noisy-student

This Kaggle notebook can be used to easily reproduce the results from this work.

Model as a Service

The ensemble can be deployed with a Docker image. The image can be retrieved here: https://hub.docker.com/r/smet/melanoma-service

The source code can be found in the /service directory.

To start the service, execute the following two commands:
docker pull smet/melanoma-service
docker run -d --name melanoma-service -p 80:80 smet/melanoma-service
The service is now available on port 80.

There are two routes. The first route returns simply a prediction. The second route returns a grad cam image.

1. /predict

Method: POST;
Parameters:

• image_url: URL of an image to predict
• number_of_models: Either 1 or 2 - How many models should be used for the prediction.
  
  

2. POST ​/cam

Method: POST
Parameters:

• image_url: URL of an image to predict

Additional documentation on how to create your own service can be found in the /docs folder.

Important Note

This application is created for scientific purposes only!