!!! Caution: click here to see application ^^ Please wait 90 seconds (max) after click ->

Mole Identification Application

Convolutional Neural Network image classifier using Keras and tensorflow backed, deployed on Heroku.

Introduction

The team has been commissioned by the health care company "skinCare" to create an application that can detect when a mole is dangerous, and advice the user to go to see the doctor.

The Team

This project is a collaborative effort between four members of the Bouwman2 promotion at BeCode, Brussels, in January 2021. The team comprises of Emre Ozan, Adam Flasse, Dilara Parry, and Naomi Thiru

Project Organization

Content	Description
Task 1	Preparation Dataset
Task 2	Importing Dataset
Task 3	Creating and Saving a Model
Task 4	Creating a Flask Application
Task 5	Deployment and Creating a Docker File

Folder Structure

Preparation of the Dataset

The downloaded data is stored in two main folders, training_data and test_data, within each, folders containing benign and malignant images. This structure is important in accessing the respective datasets and loading them for the model. This model uses ImageGenerator to augment and prepare the data for the model.

Creating and Saving a Model

Problem	Data	Methods	Libs	Link
Deep Learning model	Moles dataset	CNN	keras, tensorflow,	(https://github.com/mremreozan/challenge-mole/tree/main/app)

Flask Application

Problem	Data	Methods	Libs	Link
Deployment	Image input	GET, POST	Flask	(https://github.com/mremreozan/challenge-mole/blob/main/app/app.py)

API recieves an image file, and returns a response of whether or not one should see a doctor.

• Url:

  https://radiant-wave-54553.herokuapp.com/

• Method:

  GET POST

• Success Response:

  
  if result == 0.0(benign):
         return "Don't worry, it is not serious, this patient doesn't need to see a doctor!"
     else:
         return "This patient need to see a doctor!"```
  
  
  
  

Deployment

Problem	Data	Methods	Libs	Link
Environment		Docker	Dockerfile, requirements.txt,Procfile, tensorflow==2.3.2	(https://github.com/mremreozan/challenge-mole/blob/main/app/Dockerfile)

In case you would like to try our API and run on container on a Web Application Service, you can do this on Heroku. Using this documentation will help you to try our API with our environment prepared on Docker : https://github.com/mremreozan/challenge-mole/blob/main/app/Dockerfile

heroku login
heroku container:login
heroku create
heroku container:push web -a <heroku repository name>
heroku container:release web -a <heroku repository name>

What's left

Although the model gave an accuracy of 80%, we were not able to try out some preprocssing using openCV to see if this affects its accuracy. To further develop the CNN we could have more layers, a deeper CNN which would allow for a higher accuracy. We could also train for more epochs.

Results (first is malignant mole, second is bening mole.)