Fundus Image Optic Disc Recognition via Ellipse Fitting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning-based methods for recognizing optic disc (OD) and optic cup (OC) in fundus images face accuracy issues due to abnormalities like light leakage, rainbow spots, and lens stains, which result in multiple candidate areas and reduced detection precision.
Innovation Solution
A method involving pre-processing of fundus images, using a trained OD prediction model to identify candidate areas, performing ellipse fitting, and selecting the OD area based on vascular pixel count, along with adjusting OC candidate areas to ensure accurate positioning within the OD area, employs a device with image pre-processing, OD positioning, ellipse fitting, and OD area determining modules.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If deep learning is used to automatically learn characteristics of OD area, then the automation extent is improved, but when fundus images have abnormalities such as light leakage, rainbow spots, and lens stains, multiple candidate areas appear and measurement precision deteriorates
Solution Approach 1:
The patent segments the OD recognition process into multiple stages: initial candidate area identification by deep learning model, ellipse fitting to each candidate area, vascular pixel extraction and counting, and final selection based on vascular pixel count. This segmentation allows the system to handle multiple candidate areas by systematically evaluating each one through ellipse fitting and vascular analysis, ultimately selecting the most accurate OD area despite image abnormalities
Solution Approach 2:
The patent introduces ellipse fitting parameters and vascular pixel count as intermediary metrics between the deep learning candidate area identification and the final OD area determination. These intermediaries serve as objective criteria to distinguish the true OD area from false candidates caused by image abnormalities, enabling precise selection without manual intervention
2Adaptability or versatility
If multiple candidate areas are identified in fundus images with abnormalities, then the adaptability is improved to handle various image qualities, but the device complexity increases due to additional processing steps
Solution Approach 1:
The patent implements a self-service mechanism where the system automatically evaluates multiple candidate areas through ellipse fitting and vascular pixel analysis, then selects the correct OD area based on predefined criteria (maximum vascular pixel count). This self-service approach handles various image qualities adaptively without requiring manual intervention or complex external processing systems
Solution Approach 2:
The patent changes evaluation parameters systematically: first applying ellipse fitting to characterize candidate area shapes, then extracting vascular pixels within each ellipse, and finally selecting based on vascular pixel count. These parameter changes provide a standardized, automated method to evaluate and compare multiple candidates across different image qualities without increasing overall system complexity
Data Source
AI summary
The present disclosure relates to a method and device for recognizing a fundus image, and equipment. The method includes: obtaining an acquired fundus image, and pre-processing the acquired fundus image; inputting the pre-processed fundus image to a trained optic disc (OD) prediction model, and performing OD prediction on the fundus image by the OD prediction model to obtain a corresponding OD prediction map, where the OD prediction map is marked with one or more located candidate areas of the OD; and obtaining the OD prediction map and performing ellipse fitting to obtain ellipse fitting parameters of the candidate areas of the OD, and determining an OD area in the fundus image based on the number of the candidate areas of the OD and the ellipse fitting parameters.