Retinal Image Processing for Anatomical Feature Localization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional retinal image-processing methods face challenges in accurately and reliably determining the location of anatomical features like the optic disc and fovea in retinal images from modalities other than fundus reflectance photographs, especially due to artefacts, pathology, variations in eye position, and differences between image modalities, particularly for ultra-wide field views.
Innovation Solution
A computer-readable storage medium and image-processing apparatus that calculate local orientation vectors for blood vessels in retinal images, normalize them, and use accumulators to determine the location of anatomical features by accumulating votes from pixels offset in specific directions, with smoothing to refine the results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional methods use basic blood vessel detection algorithms, then the processing speed is fast, but the positional accuracy of anatomical features deteriorates in ultra-wide field views and non-fundus modalities
Solution Approach 1:
The patent segments the retinal image processing into distinct stages: blood vessel detection, orientation vector calculation, accumulator voting, and feature location determination. Each stage processes specific aspects independently, improving positional accuracy without overwhelming computational complexity at any single step.
Solution Approach 2:
The patent introduces orientation vectors that add directional information as a new dimension to the processing. Instead of only detecting vessel presence, the system calculates local orientation at each pixel, creating a two-dimensional feature space (position + orientation) that significantly improves anatomical feature localization accuracy.
2Measurement precision
If the method processes all pixels in the image, then the positional accuracy improves, but the computational time and resources increase
Solution Approach 1:
The patent extracts only the essential features needed for anatomical feature detection: blood vessel orientation information. By filtering out irrelevant image data and focusing computation only on orientation vectors along vessels, the system achieves high positional accuracy without processing every pixel equally, thus reducing overall computational time.
Solution Approach 2:
The patent performs preliminary blood vessel detection and orientation calculation before the accumulator voting stage. This preliminary processing identifies relevant regions and prepares orientation data in advance, so that the subsequent feature location determination can proceed efficiently with pre-processed information rather than raw image data.
3Reliability
If the system uses simple accumulator voting without smoothing, then the processing is faster, but the location determination becomes less robust against artefacts and pathology
Solution Approach 1:
The patent applies smoothing to the accumulator values after the voting process. This smoothing acts as a cushioning mechanism that reduces the impact of outliers and artefacts on the final feature location determination. By averaging nearby accumulator values, the system becomes more robust to noise and pathological variations without requiring complex filtering during the voting process itself.
Data Source
Figure 1
Figure 2
Figure 3A~3C
AI summary
There is provided a non-transitory computer-readable storage medium storing computer program instructions which, when executed by a processor, cause the processor to process image data defining an image of a retina to determine a location of an anatomical feature of the retina in the image by: receiving the image data; calculating, for each of a plurality of pixels of the received image data, a respective local orientation vector indicative of the orientation of any blood vessel present in the image at or adjacent the pixel; calculating a normalised local orientation vector for each of the plurality of pixels by normalising the local orientation vector calculated for each of the pixels so that the magnitude of the normalised local orientation vector at each of the pixels takes a common predetermined value; operating on an array of accumulators, wherein each accumulator in the array is associated with a respective pixel of the received image data, by (i) for each of the plurality of pixels, adding a respective value to an accumulator associated with a pixel of the received image data that is offset from the pixel of the plurality of pixels by the predetermined value in a predetermined direction relative to the direction of the local orientation vector at the pixel of the plurality of pixels, and (ii) smoothing the values in the accumulators; and determining the location of the anatomical feature in the image of the retina using the location of a pixel of the received image data which is associated with an accumulator having accumulated an accumulated value, which accumulated value is within a predetermined range of an extremum of the accumulated values in the accumulators.