Page 38 - AIH-1-4

Artificial Intelligence in Health Segmentation and classification of DR using CNN

images and applying decision trees, Naive Bayes, or k-NN, A B
7
yielded an accuracy of 73.4% and an F-measure of 68.4%
on a dataset of 151 images with varying resolutions. These
efforts illustrate ongoing advancements in leveraging
machine learning for DR detection.
The rise of deep learning approaches has led to the
development of various methods for applying CNNs for
DR detection. Pratt et al. developed a network with a CNN Figure 1. The original fundus image and its corresponding mask from
8
the IDRiD. This dataset, consisting of 516 images, captures lesion
architecture and data augmentation capable of identifying segmentation in fundus images taken by a retinal specialist in Nanded,
intricate features related to classification tasks such as Maharashtra, India. (A) Original fundus image and (B) the corresponding
MAs, exudates, and hemorrhages (HEs) in the retina, mask. Images obtained from the IDRiD.
providing automated diagnoses without user input. Their Abbreviation: IDRiD: Indian diabetic retinopathy image dataset.
model achieved a sensitivity of 95% and an accuracy of
75% on a validation set of 5000 images. Other researchers macula, and a resolution of 4288 × 2848 pixels stored in
have also contributed to CNN-based approaches. jpg file format, with each image approximately 800 KB in
9,10
Notably, Asiri et al. conducted a comprehensive review size. 15
11
of existing methods and datasets, highlighting their pros For assessing lesion segmentation techniques related to
and cons while emphasizing the challenges in designing DR, pixel-level annotated data includes binary masks for
efficient and robust deep learning algorithms for diverse distinct abnormalities, including MAs, hard exudates (EXs),
problems in DR diagnosis and suggesting directions for HEs, and soft exudates (SEs). The dataset comprises color
future research. fundus images in.jpg format, along with corresponding
Moreover, researchers have explored transfer learning binary masks in.tiff files. The dataset contains 81 images
using CNN architectures. Hagos and Kant attempted to with binary masks for MAs, 81 for EXs, 80 for HEs, and 40
12
train InceptionNetV3 for 5-class classification with pre- for SEs, accommodating images with multiple lesions and
training on the ImageNet dataset, achieving an accuracy enhancing robust research and performance evaluation of
of 90.9%. Sarki et al. trained ResNet50, Xception Nets, lesion segmentation techniques in DR. 36
13
DenseNets, and VGG with ImageNet pre-training,
obtaining the best accuracy of 81.3%. Both research teams 1.2.3. Image grading
utilized datasets provided by APTOS and Kaggle. Our research utilized image data sourced from multiple
datasets, primarily focusing on an open dataset obtained
1.2. Problem statement
from the Kaggle DR Detection Challenge 2015 for pre-
16
1.2.1. Datasets training CNNs. This dataset is widely recognized as
17
The image data employed in this study was obtained from the largest publicly available, comprising 35,126 fundus
diverse datasets. The research encompasses three primary photographs capturing both the left and right eyes of
objectives: lesion segmentation, disease/image grading, American citizens. The images are labeled with stages of
and treatment recommendations. DR, ranging from no DR (label 0) to proliferative DR (label
4), as illustrated in Figure 2. 18
1.2.2. Lesion segmentation In addition to the Kaggle dataset, we incorporated
Lesion segmentation in the Indian Diabetic Retinopathy other smaller datasets, including the IDRiD, as shown in
19
Image Dataset (IDRiD) uses fundus images captured by a Figure 3, from which we utilized 413 fundus photographs,
retinal specialist at an eye clinic in Nanded, Maharashtra, and the methods to evaluate segmentation and indexing
India. From the 100 of examinations available, we have techniques in the field of retinal ophthalmology
extracted 516 images to form our dataset, as shown in (MESSIDOR) dataset, contributing 1200 fundus
20
Figure 1A and B. 12 photographs. To ensure consistency, we used a version of
the MESSIDOR dataset that had been relabeled to standard
Experts verified that all images are of adequate quality 21
and clinically relevant, that no image is duplicated, and that a grading by a panel of ophthalmologists.
reasonable mixture of disease stratifications representative Evaluation of our models was conducted on the Kaggle
of DR and diabetic macular edema is present. The fundus APTOS 2019 Blindness Detection dataset, with access
22
14
images were acquired using a Kowa VX-10 alpha digital limited to the training portion. The full APTOS 2019
fundus camera with a 50° field of view, centered near the dataset comprises 18,590 fundus photographs divided into

Volume 1 Issue 4 (2024) 32 doi:10.36922/aih.2783