Unified Ocular Image Network for Classification-Led Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional deep neural networks (DNNs) used for image classification and segmentation in medical imaging require significant computational resources, large datasets for training, and face privacy concerns when deployed centrally, leading to inconsistent outcomes and inefficiencies.
Innovation Solution
A deep learning model adapted to simultaneously perform classification and segmentation, integrating classification outputs to refine segmentation processes, reducing the need for separate models and computational resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If separate deep neural networks are used for classification and segmentation, then each task can be performed independently, but the outcomes become inconsistent and inaccuracies increase
Solution Approach 1:
The patent combines classification and segmentation tasks into a single unified deep neural network model. The network architecture integrates multiple output heads that simultaneously perform classification (identifying the type of ocular disease) and segmentation (delineating affected retinal layers), allowing both tasks to benefit from shared feature extraction and consistent parameter learning, thereby improving outcome consistency and accuracy.
Solution Approach 2:
The unified neural network model serves multiple functions: it performs feature extraction, classification of ocular diseases, and segmentation of retinal layers all within a single model framework. This multi-functional approach allows the model to leverage common patterns across tasks while maintaining specialized capabilities for each function, reducing the need for separate models and improving overall reliability.
2Reliability
If a large amount of data is used to train reliable models, then model performance improves, but the computational resources and GPUs required increase significantly
Solution Approach 1:
The training process is segmented into two phases: pre-training on large public datasets to learn general features, and fine-tuning on smaller private datasets to adapt to specific applications. This segmentation allows the model to benefit from large-scale data without requiring all training data to be processed simultaneously, reducing computational resource requirements while maintaining model reliability.
Solution Approach 2:
The model undergoes pre-training on large public datasets before deployment, performing the computationally intensive learning phase in advance. This preliminary action allows the model to acquire robust features beforehand, reducing the computational burden during actual deployment and enabling reliable performance with fewer resources during operational use.
3Ease of operation
If a centralized model is deployed publicly over the internet, then accessibility improves, but privacy concerns arise
Solution Approach 1:
The system allows for localized deployment of the model at individual healthcare institutions, enabling each location to run the model locally on their own infrastructure. This approach maintains high accessibility and ease of operation while ensuring that patient data remains local and private, eliminating the need to upload sensitive medical images to centralized public servers.
Solution Approach 2:
The patent introduces a federated learning framework where the centralized model acts as an intermediary that coordinates training across multiple distributed locations without centralizing the actual patient data. Each institution trains the model locally on their private datasets, and only model parameter updates are shared, thereby maintaining both accessibility and privacy through the intermediary coordination layer.
Data Source
AI summary
A system and method for processing an image including an image gateway adapted to receive one or more input images, a learning network configured to perform feature extraction on the one or more input image, and; simultaneously perform a classification function and a segmentation function on the input image.