Neural Network Patch Segmentation for Disease Diagnosis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current disease diagnosis systems using neural networks struggle to accurately determine disease presence on biological tissues by considering adjacent tissues, leading to incomplete or inaccurate diagnoses when relying solely on patch-level analysis.
Innovation Solution
A disease diagnosis system and method that incorporates a patch-level segmentation neural network with a non-local correlation calculation sub-architecture to segment disease areas within patches, utilizing feature maps from multiple layers and non-local correlation nodes for enhanced accuracy, and a segmentation sub-architecture to produce masks for disease area identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If patch-level diagnosis is performed using conventional neural networks, then diagnosis speed is improved, but diagnosis accuracy deteriorates because adjacent tissue information is not considered
Solution Approach 1:
The system segments the biological tissue image into multiple patches and processes them through a neural network, enabling parallel processing for speed while maintaining the ability to analyze spatial relationships for accuracy
Solution Approach 2:
A non-local block is introduced as an intermediary component that captures long-range dependencies and contextual information from adjacent patches, allowing the system to maintain both processing speed and diagnostic accuracy by mediating between local patch features and global tissue context
2Measurement precision
If segmentation is performed to identify disease areas, then diagnosis accuracy is improved, but device complexity increases due to additional architecture components
Solution Approach 1:
The segmentation architecture is nested within the existing classification neural network, with feature extraction layers shared between both functions. The non-local block and segmentation head are integrated into the network hierarchy, allowing accurate disease area detection while reusing computational resources to mitigate complexity increases
3Measurement precision
If non-local blocks are added to capture adjacent tissue information, then diagnosis accuracy is improved, but computational cost increases
Solution Approach 1:
The system dynamically adjusts the scope and intensity of non-local operations based on input characteristics, modifying computational parameters to achieve high diagnostic accuracy while controlling computational cost through adaptive processing
Data Source
AI summary
A disease diagnosis system uses a slide of a biological image and the neural network, the disease diagnosis system including a patch-level segmentation neural network that receives, for each predetermined patch in which the slide is divided into a predetermined size, the patch as an input layer so as to specify the area in which the disease in the patch exists, wherein the patch-level segmentation neural network comprises: a patch-level classification neural network, which receives the patch as an input layer so as to output a patch-level classification result about whether the disease exists in the patch; and a patch-level segmentation architecture, which receives a feature map generated in each of two or more feature map extraction layers from among hidden layers included in the patch-level classification neural network, so as to specify the area in which the disease in the patch exists.