GradientAV

Gradient-Preserving Retinal Artery-Vein Classification

基于梯度保留的视网膜动静脉分类后处理方法

English | 中文

---

English

Overview

GradientAV is a post-processing tool for retinal vessel segmentation that converts neural network predictions (multi-colored: red/blue/green/purple/cyan) into standardized artery-vein classification maps (red=artery, blue=vein, black=background).

Key Innovation: Instead of the traditional "binarize-then-smooth" paradigm, we preserve the original gradient information from neural network outputs, achieving naturally smooth vessel edges without shape distortion.

Problem & Motivation

Traditional Methods	Problems
Morphological smoothing	Jagged edges remain, over-processing causes deformation
Gaussian blur + threshold	Vessels become thicker, details lost
B-spline contour interpolation	Computationally complex, prone to distortion
Super-sampling anti-aliasing	Severe structural distortion

Root Cause: All methods follow the "binarize → smooth" paradigm. Once hard edges are created, subsequent smoothing inevitably leads to information loss or deformation.

Our Approach

Core Insight: Neural network outputs inherently contain natural edge gradients (from softmax probabilities or sub-pixel responses). This gradient IS the natural anti-aliasing.

Input Image → HSV Conversion → Soft Mask Extraction → Hue Classification → Gradient Modulation → Output
                    ↓
         Preserve original brightness gradient (NO binarization)

Mathematical Formulation:

For input image $I$ with HSV components $(H, S, V)$, output image $O$:

$$O_R = \alpha(H) \cdot \frac{V}{255} \cdot 255$$

$$O_B = (1 - \alpha(H)) \cdot \frac{V}{255} \cdot 255$$

Where $\alpha(H)$ is the hue-based artery weight function with smooth boundary transitions.

Installation

git clone https://github.com/Changhuaishui/GradientAV.git
cd GradientAV
pip install -r requirements.txt

Usage

Single file:

python gradient_av.py -i input.png -o output.png

Batch processing:

python gradient_av.py -i ./input_folder -o ./output_folder

With comparison output:

python gradient_av.py -i input.png -o output.png --compare

Parameters:

Parameter	Description	Default
-i, --input	Input file or directory	Required
-o, --output	Output file or directory	Required
-c, --compare	Generate side-by-side comparison	False
-m, --min-intensity	Minimum intensity threshold	5

Results

Metric	Traditional Methods	GradientAV
Edge Smoothness	Obvious jagging	Natural gradient
Vessel Shape Preservation	Deformation/thickening	Fully preserved
Processing Speed	Slow (multiple iterations)	Fast (single mapping)

---

中文

概述

GradientAV 是一个视网膜血管分割的后处理工具，用于将神经网络预测结果（多色：红/蓝/绿/紫/青）转换为标准化的动静脉分类图（红色=动脉，蓝色=静脉，黑色=背景）。

核心创新：摒弃传统的"二值化-再平滑"范式，直接保留神经网络输出的梯度信息，实现自然平滑的血管边缘且不产生形变。

问题与动机

传统方法	问题
形态学平滑（开闭运算）	边缘仍有锯齿，过度处理导致血管变形
高斯模糊 + 阈值化	血管变粗，细节丢失
轮廓B样条插值	计算复杂，易产生形变
超采样抗锯齿	血管结构严重失真

根本原因：所有方法都遵循 "二值化→平滑" 范式。硬边缘一旦产生，后续平滑必然导致信息损失或形变。

我们的方法

核心洞察：神经网络的原始预测输出本身就包含自然的边缘渐变信息（来自 softmax 概率或网络的亚像素响应），这种渐变就是天然的抗锯齿。

输入图像 → HSV转换 → 软掩码提取 → 色相分类 → 梯度调制 → 输出
                ↓
         保留原始亮度渐变（不二值化）

安装

git clone https://github.com/Changhuaishui/GradientAV.git
cd GradientAV
pip install -r requirements.txt

使用方法

单文件处理：

python gradient_av.py -i input.png -o output.png

批量处理：

python gradient_av.py -i ./input_folder -o ./output_folder

生成对比图：

python gradient_av.py -i input.png -o output.png --compare

实验结果

指标	传统方法	GradientAV
边缘平滑度	锯齿明显	自然渐变
血管形态保持	变形/变粗	完全保持
处理速度	较慢（多次迭代）	快速（单次映射）

---

Project Structure

GradientAV/
├── README.md
├── LICENSE
├── requirements.txt
├── gradient_av.py          # Main program
├── examples/
│   ├── input/              # Example inputs
│   └── output/             # Example outputs
└── docs/
    └── method.md           # Detailed methodology

---

Star History

---

Citation

If you find this work useful, please consider citing:

@software{gradientav2024,
  author = {Changhuaishui},
  title = {GradientAV: Gradient-Preserving Retinal Artery-Vein Classification},
  year = {2024},
  url = {https://github.com/Changhuaishui/GradientAV}
}

---

License

This project is licensed under the MIT License - see the LICENSE file for details.

---

If this project helps you, please give it a star!