Diabetic-Retinopathy-Classification WebbAppBot using Deep Learning: CNN and NLP

Objective:

The objective of this project is to make information and diagnosis of Diabetic Retinopathy accessible to diabetes patients and also to assist ophthalmologist in diagnosing the disease. The project will reduce the window, between the testing and start of treatment, significantly

Table of Contents

1. Obtaining the Datasets
2. Loading and Preprocessing
   • Loading Images
   • Resize Images
   • Class Imbalance
3. Model Architecture
4. Web Deployment
5. Limitation
6. Authors

Obtaining the Datasets

• Classification Images: The scanned images originates from Kaggle competition.
• Corpus Dataset: Created the corpus by web-scraping MedicineNet.

Installation

Install:

• Anaconda
• Python (Minimum 3)
• Keras for Model deployment
• Seaborn for visualization
• OpenCV for cleaning the images
• anvil for web deployment

Loading and Preprocessing

The NLP Model was processed on the Google Colab while the Image classification was processed on my personal machine.

A database was created to keep track all the questions that the Bot was not trainned to answer. The Model is updated everyday by retraining the model

Resize All Images

The images were of different sizes. I created Height and width columns for each images and resized the images to the average of height and width using openCV before feeding the to the model

Class Imbalance

Applied data augmentation to the dataset using image data generator with keras, I

Training the Model:

After trying different method to train the classification model, ResNet50 was used for feature extraction applied on the augmented data to train the image classification model. For the natural language processing, i used fasttext to handle any out of vocabulary text using fasttext representation algoritm

Web Deployment

The webapp is hosted on anvil.works retinopathyBOT. This involve writing some Java Script with Python. Hopefully, the app is still up and running by the time you are reading this readme. This is because of the monthly payment to host the webapp on anvil.works

Limitiation

• The speech recognition API currently works of firefox and chrome browser
• Cloud deployment cost money, so the app might not be on for long time

Authors

Omolewa Davids

• website
• LinkedIn