Neural Network Avascular Map Generation from OCTA Artifacts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current medical imaging technologies, particularly Optical Coherence Tomography Angiography (OCTA), face challenges in accurately detecting avascular areas in retinas due to artifacts and noise, which hinders the diagnosis and quantification of Non-Perfusion Areas (NPAs) in Diabetic Retinopathy (DR) patients.
Innovation Solution
A neural network system, MEDnet-V2, is developed to generate avascular maps from OCT images by merging reflectance intensity maps, thickness images, and OCTA images, using a multi-scaled encoder-decoder architecture that distinguishes between vascular, avascular, and signal reduction areas, effectively reducing the impact of artifacts and noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If OCTA images are used to detect avascular areas, then high resolution and depth resolution are achieved, but artifacts and noise confound the interpretation of data
Solution Approach 1:
The patent introduces a neural network as an intermediary processing layer between the OCTA images and the final avascular area detection. This neural network mediator learns to distinguish true avascular areas from artifacts by training on annotated data, effectively filtering out harmful factors while preserving useful information.
Solution Approach 2:
The patent replaces traditional image processing mechanisms with a neural network-based system. Instead of using conventional algorithms that are susceptible to artifacts, the neural network uses learned patterns from training data to robustly identify avascular areas, substituting mechanical image processing with intelligent pattern recognition.
2Measurement precision
If traditional image processing methods are used, then the system is simpler, but accuracy in distinguishing true NPAs from artifacts is insufficient
Solution Approach 1:
The patent performs preliminary action by training the neural network on annotated OCTA images before actual detection. During training, the network learns to distinguish between true avascular areas and artifacts, preparing it to accurately detect NPAs in new images without requiring complex real-time processing during diagnosis.
Solution Approach 2:
The patent changes the fundamental parameter of image processing from traditional algorithms to neural network-based pattern recognition. This parameter change enables the system to achieve high accuracy in distinguishing true NPAs from artifacts by learning complex patterns from training data, accepting the resulting system complexity as necessary for improved performance.
3Measurement precision
If multi-scaled encoder-decoder architecture is used, then distinction between vascular, avascular, and signal reduction areas is improved, but processing time increases
Solution Approach 1:
The patent applies segmentation by dividing the retinal area classification task into distinct segments: vascular areas, avascular areas, and signal reduction areas. The multi-scaled encoder-decoder architecture processes different scales of the input image to simultaneously identify all three types of regions, improving classification accuracy through systematic division of the detection task.
Data Source
AI summary
This disclosure describes systems, devices, and techniques for training neural networks to identify avascular and signal reduction areas of Optical Coherence Tomography Angiography (OCTA) images and for using trained neural networks. By identifying signal reduction areas in OCTA images, the avascular areas can be detected with high accuracy, even when the OCTA images include artifacts and other types of noise. Accordingly, various implementations described herein can accurately identify avascular areas from real-world clinical OCTA images. In various implementations, a method can include identifying images of retinas. The images may include thickness images, reflectance intensity maps, and OCTA images of the retinas. Avascular maps corresponding to the OCTA images can be identified. A neural network can be trained based on the images and the avascular maps.