Multi-Resolution Tissue Image Segmentation for Spatial Continuity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Processing large gigapixel images in digital pathology is impractical due to computational and memory limitations, leading to information loss when high-resolution images are divided into smaller patches, which disrupts the continuity and overall structure of the image.
Innovation Solution
A multi-resolution image segmentation approach using a set of images with varying resolutions, where each encoder-decoder is trained separately to process specific resolutions, and feature information is passed between encoder-decoders to maintain semantic relationships and improve processing efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If high-resolution gigapixel images are divided into smaller patches for processing, then computational and memory limitations are addressed, but spatial continuity and overall image structure are disrupted and information is lost
Solution Approach 1:
The patent applies segmentation by dividing the gigapixel image into multiple context images at different resolutions (e.g., 200x200, 500x500, 1000x1000 pixels). Each context image covers the same tissue area but at a different scale, allowing the system to process manageable image sizes while preserving spatial relationships through the hierarchical resolution structure.
Solution Approach 2:
The patent implements a nested structure where context images at different resolutions are hierarchically organized. Lower-resolution context images provide broad spatial context that is progressively refined in higher-resolution context images, similar to nested dolls. This nesting preserves spatial continuity information across scales while enabling efficient processing of each individual context image.
2Productivity
If multiple context images at different resolutions are processed separately through encoder-decoders, then processing efficiency is improved, but system complexity increases
Solution Approach 1:
The system segments the processing task by training separate encoder-decoder models for each context image resolution. This segmentation allows parallel processing of different resolution contexts independently, improving processing efficiency while managing complexity through modular model design where each encoder-decoder is optimized for its specific resolution input.
Solution Approach 2:
The encoder-decoder architecture serves multiple functions across different resolutions. The same fundamental encoder-decoder model structure is universally applied to process context images at various resolutions (200x200, 500x500, 1000x1000), allowing a single model design to handle multiple processing tasks and reducing overall system complexity through reuse.
3Measurement precision
If feature information is aggregated from multiple resolutions to determine pixel classification, then diagnostic accuracy is enhanced, but computational overhead increases
Solution Approach 1:
The patent applies local quality by using higher-resolution context images for detailed pixel classification decisions while relying on lower-resolution context images for broader spatial context. Each resolution level contributes features appropriate to its scale, with finer details from high-resolution images and general patterns from low-resolution images, optimizing the balance between classification accuracy and computational energy usage.
Solution Approach 2:
The system performs preliminary processing by generating context images at multiple resolutions and extracting features from each resolution level before final pixel classification. This preliminary action at different scales prepares feature information in advance, allowing the final classification step to aggregate pre-processed features efficiently, reducing the computational energy required during the actual classification process.
Data Source
AI summary
Systems and methods for determining pixel classification information using images depicting at least a portion of a whole slide image (WSI) of a stained tissue sample. A system can store a first image of the tissue sample at a first resolution, a second image of the tissue sample at a second resolution that is higher than the first resolution, and a third image of the tissue sample at a third resolution that is higher than the second resolution, the first, second, and third images depicting at least a portion of a same area of the tissue sample. The system can include be configured to generate first feature information based on the first image, generate second feature information based on the second image, and determine pixel classification information of at least a portion of the WSI based on the third image, the first feature information and second feature information.