Choroid-Scleral OCT Segmentation With DCT Decoders and Uncertainty Refinement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI architectures for retinal layer segmentation in OCT data face challenges in the training process, which is resource-intensive and time-consuming, making revisions and retraining difficult.
Innovation Solution
A deep-learning architecture for choroid-scleral layer segmentation using a neural network with a contracting path and an expanding path that includes discrete cosine transform decoders, omitting deconvolution modules, and an uncertainty detection module to refine initial segmentations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional CNN architectures are used for retinal layer segmentation, then segmentation accuracy can be achieved, but the training process becomes resource-intensive and time-consuming
Solution Approach 1:
The patent extracts and removes the deconvolution module from the traditional CNN architecture, retaining only the contracting path with convolution layers and max pooling layers. This extraction simplifies the network structure, reduces the number of parameters, and significantly decreases training time and computational resource requirements while maintaining segmentation accuracy through the uncertainty detection module.
2Measurement precision
If traditional CNN architectures are used for retinal layer segmentation, then segmentation accuracy can be achieved, but the model complexity and number of parameters increase
Solution Approach 1:
The patent removes the deconvolution module from the traditional U-Net or VGG-based CNN architectures, keeping only the contracting path consisting of convolution layers and max pooling layers. This extraction dramatically reduces the number of model parameters and simplifies the overall architecture, making the model more efficient and easier to train while maintaining segmentation performance through alternative mechanisms.
Solution Approach 2:
Instead of using deconvolution operations to upsample feature maps in the expanding path, the patent inverts the approach by using the uncertainty detection module to identify and correct segmentation errors in the reduced contracting path output. This inversion of the traditional decoder architecture achieves accurate segmentation with a simpler, more parameter-efficient model.
3Productivity
If traditional CNN architectures are used for retinal layer segmentation, then segmentation can be performed, but revisions and retraining become difficult
Solution Approach 1:
By extracting and removing the deconvolution module, the patent creates a simpler model with fewer parameters and a more straightforward architecture consisting of only convolution and pooling operations. This simplification makes the model easier to train, modify, and retrain, as there are fewer complex components to configure and tune, thereby facilitating easier revisions and retraining processes.
4Loss of time
If a simplified neural network architecture is used, then training resources and time are reduced, but segmentation accuracy may be compromised
Solution Approach 1:
The patent introduces an uncertainty detection module that acts as a feedback mechanism, examining the segmentation output from the simplified contracting path and identifying regions of uncertainty or incorrect segmentation. This feedback loop allows the model to correct errors and maintain high segmentation accuracy despite the reduced architectural complexity and fewer parameters, effectively compensating for the simplified structure.
Data Source
AI summary
A System/Method/Device for segmenting the choroid-scleral layer from an optical coherent tomography (OCT) volume scan. The present system uses a deep learning machine model based on a neural network that include multiple convolution layers, but no deconvolution layers. Rather, the present neural network is based on a novel architecture based on the discrete cosine transform.