Eye Disease Recognition - AIMedic

Contributers

Name
Mahsa Amani
Afrooz Shekholeslami
Ali Ghadiri
Amin Jamshidi
Mojtaba Khanloo
Asiyeh Bahaloo

Abstraction

As retinal pathologies are becoming more and more common globally, rapid and accuract detecting of eye diseases plays a crucial role in preventing from blindness. Therefore deep learning approaches will sharply increase probability of curing such diseases.

Introduction

According to World Health Organization at least 2.2 billion people have a near or distance vision impairment. In at least 1 billion – or almost half – of these cases, vision impairment could have been prevented or has yet to be addressed.

This 1 billion people includes those with moderate or severe distance vision impairment or blindness due to unaddressed refractive error (88.4 million), cataract (94 million), glaucoma (7.7 million), corneal opacities (4.2 million), diabetic retinopathy (3.9 million), and trachoma (2 million), as well as near vision impairment caused by unaddressed presbyopia (826 million). Population growth and ageing are expected to increase the risk that more people acquire vision impairment.

In the past few years , deep learning has helped lots of people to live better around the world. In this case we will use deep learning to detect different eye pathologies. Some critical articles in this area were studied. Some articles used preprocessing methods to improve the image. Then they trained several models to get features and finaly they predicted the potential disease. Some other articles developed models to detect some lesions such as microaneurysms, hemorrhages, exudates, and cotton-wool spots.

Datasets

We have used two datasets to train our models:

1. ODIR_2019 dataset (downlaod):
   This dataset is real-life set of patient information collected by Shanggong Medical Technology Co., Ltd. from different hospitals/medical centers in China. In these institutions, fundus images are captured by various cameras in the market, such as Canon, Zeiss and Kowa, resulting into varied image resolutions.

2. Cataract dataset (downlaod): Cataract and normal eye image dataset for cataract detection.

After exploring datasets, we used some preprocessing techniques to improve images:

1. Resize images : We resized images to 224*224 pixels
2. Remove padding : We removed the padding around the fundas image and cropped uninformative area to detect lesions better.
3. Ben_Graham Method : Ben Graham (Kaggle competition's winner) share insightful way to improve lighting condition. Here, we applied his idea, and we could see many important details in the eyes much better.

Methodology

Firstly we explored datasets. Then we preprocessed datasets with methods mentioned in Dataset section. Next we trained five different deep learning models to classify images into 8 groups to extract relevant image features and automatic detection of eye diseases in fundas photographs. Our models:

1. VGG16
2. VGG19
3. Resnet_V2
4. Inception_V3
5. Xception

Finally, different results from the experiments were generated and compared using MLflow.

Results

In this section we will show and compare the results of our models.

• VGG16 We runned VGG16 model for 50 epochs with a batch size of four. Training data had 5896 images and validation data included 655 validation images. Also there was 1000 images for testing VGG16 model. Here are the results on test data: Training:

accuracy	auc	loss	precision	recall
0.8939	0.9061	0.2444	0.6380	0.4716

Validation:

accuracy	auc	loss	precision	recall
0.8326	0.7990	0.3813	0.3757	0.2374

• VGG19 We runned VGG19 model for 50 epochs with a batch size of four. Training data had 5896 images and validation data included 655 validation images. Also there was 1000 images for testing VGG19 model. Here are the results on test data: Training:

accuracy	auc	loss	precision	recall
0.8910	0.9046	0.2492	0.6201	0.4692

Validation:

accuracy	auc	loss	precision	recall
0.8498	0.7942	0.4214	0.4527	0.1833

• Resnet_V2

  We runned Resnet_V2 model for 100 epochs with a batch size of four. Training data had 5896 images and validation data included 655 validation images. Also there was 1000 images for testing Resnet_V2 model. Here are the results on test data:

  Training:

accuracy	auc	loss	precision	recall
0.9511	0.9771	0.1249	0.8721	0.7449

Validation:

accuracy	auc	loss	precision	recall
0.8403	0.7779	0.5965	0.3778	0.3559

• Inception_V3

  We runned Inception_V3 model for 100 epochs with a batch size of four. Training data had 5896 images and validation data included 655 validation images. Also there was 1000 images for testing Inception_V3 model. Here are the results on test data:

  Training:

accuracy	auc	loss	precision	recall
0.8660	0.7993	0.3166	0.4628	0.0337

Validation:

accuracy	auc	loss	precision	recall
0.8553	0.5603	16.198	0	0

• Xception We runned Xception model for 100 epochs with a batch size of four. Training data had 5896 images and validation data included 655 validation images. Also there was 1000 images for testing Xception model. Here are the results on test data:

  Training:

accuracy	auc	loss	precision	recall
0.9320	0.9566	0.1691	0.8190	0.6289

Validation:

accuracy	auc	loss	precision	recall
0.8660	0.8435	0.3380	0.5469	0.3535

• All in One

  Training:

  auc plot:

  precision plot:
  

  recall plot:

  Validation:

  auc plot:

  precision plot:
  

  recall plot:

Conclusion

• This project studied five deep learning models for the multiple classification of diseases.
• We faced several challenges due to the initial data imbalance.