Eye Image Generation Using GANs for OCT-UBM Modality Gaps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing imaging technologies like OCT and UBM have limitations in capturing comprehensive eye structures due to iris obstruction and resolution trade-offs, limiting their effectiveness in ophthalmic procedures.
Innovation Solution
A method and system utilizing a Generative Adversarial Network (GAN) to transform images between different types, such as OCT and UBM, for improved resolution, distortion correction, and generating postoperative images, enabling comprehensive eye structure visualization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If OCT imaging is used to capture eye structures, then high-resolution images are obtained, but the illuminating beam cannot penetrate across the iris leaving peripheral lens blocked
Solution Approach 1:
The patent uses a generative adversarial network (GAN) to create synthetic UBM images from OCT images. The GAN learns the mapping between OCT and UBM image characteristics and generates virtual UBM images that replicate the penetration capability through the iris while maintaining the high resolution of the original OCT images. This copying approach allows the system to obtain images with both high resolution and complete iris penetration without requiring actual UBM imaging.
2Adaptability or versatility
If UBM imaging is used to penetrate the iris, then complete eye structure visualization is achieved, but resolution and penetration depth trade-off occurs
Solution Approach 1:
The patent employs a GAN-based system where a first GAN network transforms OCT images into synthetic UBM images, and a second GAN network transforms these synthetic UBM images back into enhanced OCT images. This copying mechanism allows the system to leverage the iris penetration capability of UBM while preserving and enhancing the high resolution characteristics of OCT images, effectively resolving the resolution-penetration trade-off.
Solution Approach 2:
The patent introduces synthetic UBM images as an intermediary representation between actual OCT images and the desired high-resolution penetrating views. The GAN networks learn to translate between these representations, creating a virtual pathway that combines the advantages of both imaging modalities without requiring actual UBM acquisition.
3Adaptability or versatility
If multiple imaging modalities are integrated, then comprehensive eye structure visualization is achieved, but system complexity increases
Solution Approach 1:
Instead of integrating multiple physical imaging devices, the patent uses GAN-based image synthesis to create virtual representations of different imaging modalities from a single input modality. The system takes OCT images and generates synthetic UBM images and vice versa, achieving multi-modality coverage through computational means rather than physical integration, thereby reducing hardware complexity.
Solution Approach 2:
The patent replaces the mechanical complexity of integrating multiple imaging devices with a computational system based on deep learning. The GAN networks perform the function of modality transformation through software algorithms, substituting the need for physical integration of OCT and UBM devices while achieving comprehensive imaging coverage.
Data Source
AI summary
A method of augmenting an eye image includes obtaining a first training image of a first type and applying a first function to obtain a first generated image of a second type. A second function is applied to the first generated image to obtain a second generated image of the first type. A first loss function compares the first training image to the second generated image. The method also includes obtaining a second training image of the second type and applying the second function to obtain a first generated image of the first type. The first function is applied to the first generated image to obtain a second generated image having the second type. A second loss function compares the second training image to the second generated image. At least one of the first or second function are updated based on an output of the first and second loss functions.