Fundus Image Analysis Using Deep Learning Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for diagnosing glaucoma are costly and inconvenient, as they rely on optical coherence tomography, which is expensive and difficult to obtain, while color fundus images are more accessible but require effective interpretation tools.
Innovation Solution
An image processing system and method that calculates the cup-to-disc ratio and probability of optic nerve defects in color fundus images using a U-Net deep learning segmentation network, assisting doctors in glaucoma judgment by integrating this information to assess glaucoma severity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If optical coherence tomography is used to diagnose glaucoma, then diagnostic accuracy is improved, but cost and equipment accessibility deteriorate
Solution Approach 1:
The patent uses color fundus images as a substitute (copy) for optical coherence tomography images. Instead of requiring the expensive and difficult-to-obtain OCT equipment, the system processes readily available color fundus photographs to extract glaucoma diagnostic information, thereby maintaining diagnostic capability while improving equipment accessibility
Solution Approach 2:
The patent replaces the expensive, complex OCT imaging system with a low-cost deep learning-based image processing system that works with standard color fundus cameras. This substitution uses inexpensive computational resources and software algorithms instead of costly medical imaging hardware
2Measurement precision
If deep learning segmentation networks are used to analyze fundus images, then segmentation precision is improved, but computational complexity increases
Solution Approach 1:
The patent divides the complex task of glaucoma diagnosis into multiple segmentation stages: first segmenting the optic disc, then segmenting the optic cup within the disc, and finally calculating the cup-to-disc ratio. This multi-stage segmentation approach breaks down the complex computational problem into manageable steps, improving both precision and computational efficiency
Solution Approach 2:
The patent performs preliminary preprocessing of the fundus image before deep learning segmentation, including color normalization, contrast enhancement, and optic disc localization. These preliminary actions prepare the image data to reduce the complexity burden on the subsequent deep learning segmentation networks, enabling them to focus on precise boundary detection
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An image processing system includes an ophthalmoscope device and a processor. The ophthalmoscope device is configured to obtain a color fundus image. The processor is configured to receive the color fundus image; generate a blood vessel segmentation image that corresponds to the color fundus image using a computer vision algorithm or a deep learning model; preprocess the color fundus image and the blood vessel segmentation image to obtain an initial input image; and input the initial input image into a convolutional neural network. The convolutional neural network outputs a value. In addition, the processor generates fundus image analysis information from the cup-to-disc ratio and the value.