Classification of Diabetic Retinopathy using Pre-Trained Deep Learning Models

Introduction

Here is a summary of the provided section:

Diabetic Retinopathy (DR) is a leading cause of blindness worldwide. It results from clogging of the eye's blood vessels, causing two main reactions. The first reaction is the creation of new blood vessels in the vitreous area, obstructing the path of light that needs to reach the retina for vision. The second reaction involves blood leakage from vessels which damages the macula, the part of the retina responsible for detailed central vision.

The severity of diabetic retinopathy is classified into five levels as shown in Figure 1. The paper provides background on how DR affects the eye's anatomy and visual function. Proper blood flow to the retina is critical for converting light signals into neural impulses that are transmitted to the brain for image processing and sight.

The study focused on developing a machine learning application for multi-categorical classification of diabetic retinopathy (DR) using convolutional neural networks (CNN) and four pre-trained deep learning models: VGG16, MobileNet, InceptionV3, and InceptionResNetV2. The researchers utilized 1000 color fundus images from the KAGGLE diabetic retinopathy dataset, with resolutions of 350x350 and 224x224 pixels. Several image augmentation techniques were applied before feeding the images to the pre-trained models.

PREVIOUS WORK

The paper summarizes several research works on classifying diabetic retinopathy (DR) stages from color fundus images using deep learning techniques. The key points are:

Shorav Suriyal et al. classified images into two classes (with/without DR) using a MobileNet-inspired network with 25 convolutional layers. After preprocessing the 16,798 Kaggle images, they achieved 73.3% accuracy.

Arkadiusz Kwasigroch et al. classified 88,000 images into five DR stages using a VGG-D based transfer learning model. After preprocessing steps like resizing and normalization, their model with convolutional, pooling, dropout and fully connected layers achieved 81.7% accuracy using a specialized labeling technique.

Xiaoliang Wang et al. classified 166 Kaggle images into five DR stages using transfer learning on AlexNet, VGG-16 and Inception-V3 architectures after resizing images. Inception-V3 with fine-tuned optimization achieved the highest accuracy of 63.23%.

Safaraz Masood et al. classified the Eye-Paces Kaggle dataset into five classes using Inception-V3 transfer learning. After preprocessing like resizing and highlighting blue channels, their best model achieved 48.2% test accuracy.

Ardianto et al. proposed Deep-DR-Net to classify 237 images into three stages using convolutional feature extraction, coding, and classification stages inspired by ResNet and InceptionNet. With the FINdERS dataset, they achieved 60.82% accuracy.

The Data Set

The researchers utilized a dataset from the KAGGLE competition, comprising 35,126 color fundus images captured from different eye positions (left and right). The images were scaled between 0 and 4 to correspond to the five stages of diabetic retinopathy (DR) disease: Normal, Mild, Moderate, Severe, and Proliferative Diabetic Retinopathy (PDR).

V PROPOSED WORK AND METHODOLOGY

The paper discusses the implementation of a method for diabetic retinopathy classification using convolutional neural networks (CNNs) on the KAGGLE dataset of 1000 color fundus images. The key points are:

Image Preprocessing:

• Images were resized to 350x350x3 and 224x224x3 resolutions
• Data was split into 80% training, 10% validation, and 10% testing sets
• Images were standardized to [0,1] range
• Five disease classes were defined with one-hot encoded vectors

Data Augmentation:

• Various geometric transforms were applied to increase the dataset size to 2000 images and reduce overfitting

CNN Architecture:

• A custom CNN was built with convolutional, batch normalization, max pooling, dropout, flatten, and dense layers
• The final dense layer had 5 neurons for classification into the 5 disease classes

Transfer Learning:

• Pre-trained models (VGG16, MobileNet, InceptionV3, InceptionResNetV2) were fine-tuned by freezing lower layers and adding custom layers
• This leveraged features learned on ImageNet while fine-tuning for the diabetic retinopathy task

Web Application:

• The best performing model was deployed using Flask to build a web application for image classification

Evaluation Metrics:

• Area under the ROC curve (AUC) was used as the key metric, suitable for the imbalanced dataset
• Performance of the models is reported, with InceptionResNetV2 achieving the highest AUC of 0.69

Conclusions and Future Work

The paper evaluates the performance of different pre-trained deep learning models for classifying diabetic retinopathy disease using a Kaggle dataset of retinal images. The models explored include CNNs built from scratch, MobileNet, VGG16, InceptionV3, and InceptionResNetV2.

CNNs built from scratch performed poorly due to limited training data and shallow network depth. MobileNet, with its depth-wise separable convolutions, showed better performance than VGG16 which had fewer layers.

InceptionV3 and InceptionResNetV2, with their deeper architectures and ability to handle varying resolutions, demonstrated state-of-the-art performance on this medical image classification task. Fine-tuning the pre-trained InceptionV3 model with techniques like image augmentation, layer-wise fine-tuning, and adding custom layers improved the AUC metric to 0.63 compared to 0.59 reported in previous work.

InceptionResNetV2 exhibited the best overall performance due to its extremely deep architecture of 572 layers. Future work includes using cloud environments for training on larger datasets, higher resolutions, mobile app development, and integration with camera for real-time diabetic retinopathy screening.

The key advantages of transfer learning with deep pre-trained models like Inception were their ability to leverage large initial datasets like ImageNet, handle varying resolutions, and extract relevant features from the medical images despite domain shift.