Probability distribution guided optic disc and cup segmentation from fundus images.

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Deep learning models face uncertainty due to variations in sensors, data samples, and labeling accuracy.
• Accurate segmentation of optic disc (OD) and optic cup (OC) is crucial for diagnosing glaucoma and other optic nerve conditions.

Purpose of the Study:

• To propose and validate a probability distribution guided network for robust segmentation of OD and OC in fundus images.
• To address inherent uncertainties in deep learning for medical image segmentation.

Main Methods:

• A variational autoencoder (VAE) based network was developed to estimate the joint distribution of input images and segmentation masks.
• A Dilated Inception Block (DIB) was designed for improved model generalization and multi-scale feature extraction.
• The proposed network integrates pixel-wise information with semantic probability distributions.

Main Results:

• The proposed method demonstrated superior segmentation performance compared to state-of-the-art approaches on the ORIGA and REFUGE datasets.
• Achieved high mean Dice overlap coefficients: 96.57% (OD) and 88.46% (OC) on ORIGA, and 95.81% (OD) and 88.91% (OC) on REFUGE.
• The VAE-based approach effectively learned from data distributions, improving segmentation accuracy.

Conclusions:

• The probability distribution guided network offers a significant advancement in optic disc and optic cup segmentation.
• This approach enhances the reliability of deep learning models in medical image analysis by accounting for data uncertainty.
• The Dilated Inception Block contributes to better feature extraction and model generalization for improved segmentation outcomes.