Retinal Vessel Segmentation

A deep learning-based approach for automatic segmentation of blood vessels in retinal fundus images using an EfficientNetB7-UNET architecture.

Table of Contents

• Overview
• Key Features
• Model Architecture
• Dataset
• Evaluation Metrics
• Requirements

Overview

This project implements an advanced deep learning model for retinal vessel segmentation, a critical task in the diagnosis and treatment of various ophthalmological diseases. The system uses a modified U-NET architecture with an EfficientNetB7 encoder to accurately segment blood vessels from retinal fundus images.

Retinal vessel segmentation plays a vital role in the early detection and monitoring of conditions such as diabetic retinopathy, glaucoma, and age-related macular degeneration. Automatic segmentation tools help ophthalmologists make faster and more accurate diagnoses, potentially improving patient outcomes.

Key Features

• EfficientNetB7-UNET Architecture: Combines the powerful feature extraction capabilities of EfficientNetB7 with the precise localization abilities of UNET.
• High Accuracy: Achieves state-of-the-art performance on standard retinal vessel segmentation benchmarks.
• GPU Acceleration: Optimized for training and inference on NVIDIA GPUs.
• Pre-trained Models: Includes pre-trained weights for immediate use.
• Data Augmentation: Implements various augmentation techniques to improve model generalization.
• Evaluation Metrics: Provides comprehensive metrics including accuracy, precision, recall, F1-score, IoU, and AUC.
• Visualization Tools: Includes utilities to visualize the segmentation results and training progress.

Model Architecture

The architecture combines EfficientNetB7 as an encoder (feature extractor) with a UNET-style decoder for precise segmentation:

1. Encoder: Pre-trained EfficientNetB7 backbone, which provides powerful feature extraction while being computationally efficient.
2. Skip Connections: Connect encoder layers to decoder layers, preserving spatial information.
3. Decoder: A series of upsampling blocks that gradually restore the spatial dimensions while decreasing feature channels.
4. Output Layer: A 1x1 convolution with sigmoid activation to produce the final segmentation mask.

The model's effectiveness comes from:

• Transfer learning using pre-trained weights
• Skip connections to preserve spatial information
• Attention mechanisms to focus on relevant features
• Custom loss function combining binary cross-entropy and Dice loss

Dataset

This project uses the Retina Segmentation Dataset available on Kaggle, which includes retinal fundus images and their corresponding vessel segmentation masks.

Evaluation Metrics

The following metrics are used to evaluate the model:

• Accuracy: Overall pixel-wise accuracy
• Sensitivity (Recall): Ability to detect vessel pixels
• Specificity: Ability to identify non-vessel pixels
• F1-Score: Harmonic mean of precision and recall
• Area Under ROC Curve (AUC): Model's ability to distinguish between vessel and non-vessel pixels
• Intersection over Union (IoU): Overlap between predicted and ground truth masks

Requirements

• Python 3.7+
• CUDA-capable GPU (recommended for training)
• TensorFlow 2.x
• OpenCV
• scikit-learn
• scikit-image
• pandas
• numpy
• matplotlib
• albumentations