The region of interest localization for glaucoma analysis from retinal fundus image using deep learning

Area of Science:

• Medical Imaging
• Computer Vision
• Ophthalmology

Background:

• Automated retinal fundus image analysis is crucial for early detection of eye diseases like glaucoma and diabetic retinopathy.
• Accurate localization of the Optic Disc (OD) as a Region of Interest (ROI) is vital for disease diagnosis.

Purpose of the Study:

• To develop an automated method for detecting the Optic Disc (OD) bounding box in retinal fundus images.
• To reframe ROI detection as a regression problem solvable with Convolutional Neural Networks (CNNs).

Main Methods:

• A CNN was trained on public datasets (MESSIDOR, Kaggle) using data augmentation.
• Images were processed using a 13x13 grid, with each cell predicting bounding boxes and class probabilities.
• K-means clustering initialized network anchors, and a sum-squared loss function was used during training, followed by Non-maximum suppression.

Main Results:

• The proposed method achieved high accuracy: 99.05% on Kaggle and 98.78% on MESSIDOR test sets for ROI detection.
• The network detected ROIs in fundus images with a latency of 0.0045 seconds (25 ms).
• The approach demonstrated superior performance compared to recent methods without relying on handcrafted features.

Conclusions:

• The CNN model accurately detects ROIs in retinal images, including low-quality ones, without bias.
• This novel method provides a faster and more reliable approach for diagnosing eye diseases.
• The network offers a more comprehensive depiction compared to previous methodologies.