BCNN-for-Ocular-Disease-Classification

Adaptive Computation and Machine Learning Project: Bayesian Convolutional Neural Network for Detection of Cataract Ocular Disease

Project Overview

• Built a Bayesian Convolutional Neural Network (BCNN) to classify Cataract Images of patients
• A Standard Convolutional Neural Network was built as a benchmark model for performance comparison with the BCNN model
• Kaggle Dataset: Ocular Disease Recognition
• Evaluation metric: Accuracy
• Conducted two experiments to observe the generalization of the BCNN model
• Bayesian CNN Generalization Experiment 1: Tuning the number of neurons
• Bayesian CNN Generalization Experiment 2: Tuning the learning rate

Workflow

Context

Ocular Disease: Ocular diseases are those that affect the eye health and vision in patients of all different ages.

Cataract: A cataract is a dense, cloudy area that forms in the lens of the eye. A cataract begins when proteins in the eye form clumps that prevent the lens from sending clear images to the retina. The retina works by converting the light that comes through the lens into signals. Cataract is one of the most prevalent ailments that can lead to blindness.

Why Bayesian Convolution Neural Network?

• Cataract diagnosis is not a simple classification/detection problem
• Uncertainty quantification plays an important role in model classifications due to the risks/stakes involved in incorrectly classifying a patient as not having cataract.
• BCNN provides an opportunity to reason under uncertainty which is crucial in medical practice.
• BCNN enables the capturing of Aleatoric and Epistemic uncertainty
• Aleatoric uncertainty is the uncertainty that arises from the quality of the data
• Epistemic uncertainty is the uncertainty that arises from the model itself

Resources

Python Version: 3.9.7
Main Python Packages: numpy, pandas, sklearn, cv2, tensorflow, keras, seaborn, matplotlib, random, pickle
Python Packages: See requirements.txt file
Kaggle Dataset: https://www.kaggle.com/datasets/andrewmvd/ocular-disease-recognition-odir5k
Install prerequisite libraries:

pip install -r requirements.txt

BCNN Model Performance

• Epochs: 100, Learning rate: 0.001, Accuracy: 93.16%, Validation Accuracy: 92.50%

BCNN Model Classification/Prediction Samples

Cataract Patient Classification


Cataract Patient Classification


Normal Patient Classification


Normal Patient Classification

Standard CNN (Left) and Bayesian CNN (Right) Network Architecture

• The constructed Neural Ntework used in the project for the Standard CNN and Bayesian CNN