A deep convolutional neural network-based novel class balancing for imbalance data segmentation.

Area of Science:

• Ophthalmology and Medical Imaging
• Computer Vision and Machine Learning

Background:

• Retinal fundus images are vital for diagnosing eye conditions, revealing structures like blood vessels, optic disk, macula, and fovea.
• Accurate segmentation of retinal blood vessels is hindered by imbalanced pixel distribution and variations in vessel thickness.

Purpose of the Study:

• To propose a novel deep learning pipeline, BLCB-CNN, for accurate retinal blood vessel segmentation.
• To address challenges of data imbalance and varying vessel thickness in retinal fundus images.

Main Methods:

• Developed a Bi-Level Class Balancing Convolutional Neural Network (BLCB-CNN) incorporating Level-I (vessel/non-vessel) and Level-II (thick/thin vessel) balancing.
• Employed pre-processing techniques including Global Contrast Normalization (GCN), Contrast Limited Adaptive Histogram Equalization (CLAHE), and gamma corrections.
• Utilized a classification-based segmentation approach on a balanced dataset derived from pre-processed retinal images.

Main Results:

• Achieved superior performance on standard retinal fundus images with an Area Under the ROC Curve of 98.23%, Accuracy of 96.22%, Sensitivity of 81.57%, and Specificity of 97.65%.
• Demonstrated strong generalization ability through external cross-validation on STARE images.

Conclusions:

• The BLCB-CNN pipeline effectively segments retinal blood vessels by addressing data imbalance and enhancing image quality.
• The proposed method shows significant potential for improving the analysis of retinal vascular structures in clinical settings.