Digital Tissue Segmentation via Overlapping Tile Aggregation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current medical image analysis techniques, particularly for large whole-slide histopathology images, face challenges in processing and interpreting due to their size, requiring decomposition into smaller tiles and lack of transparency in classification rationale, which complicates clinical adoption and accurate tissue segmentation.
Innovation Solution
The method involves decomposing digital images into overlapping subimages based on visual criteria like image entropy, using classifiers like CNNs or autoencoders to generate classification probabilities, and aggregating these to achieve accurate tissue segmentation without prior knowledge of tissue types, enabling characterization of diseased tissue and enhancing classification accuracy through data augmentation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If large whole-slide histopathology images are decomposed into smaller tiles for CNN processing, then the images become amenable to computational analysis, but the classification rationale becomes opaque and tissue segmentation accuracy deteriorates
Solution Approach 1:
The patent divides large whole-slide images into smaller overlapping tiles that can be processed by CNNs. Each tile is classified independently, and the results are aggregated to produce the final segmentation map. This segmentation approach maintains computational feasibility while preserving tissue segmentation accuracy through overlapping regions that provide multiple views of the same area.
Solution Approach 2:
The patent introduces a new dimension by creating overlapping tile representations of the same image space. Instead of processing each non-overlapping region once, the same spatial location appears in multiple overlapping tiles, allowing the system to aggregate classifications from multiple perspectives to improve segmentation accuracy and provide classification rationale.
2Measurement precision
If traditional segmentation techniques are used to partition images into tissue classes, then tissue segmentation can be achieved, but the techniques become complex and computationally demanding
Solution Approach 1:
The patent uses CNNs to learn tissue segmentation patterns from training data, creating a computational model that copies and generalizes segmentation knowledge. Once trained, the CNN can perform segmentation without requiring complex hand-crafted features or domain expertise, simplifying the technique while maintaining or improving accuracy.
Solution Approach 2:
The patent transforms the segmentation problem from a complex manual or rule-based process into a parameter-optimized machine learning problem. By adjusting CNN architecture parameters, learning rates, and training data composition, the system achieves accurate segmentation with a unified framework that is less complex than traditional multi-step segmentation pipelines.
3Area of stationary object
If CNNs process every pixel in the image, then complete coverage is achieved, but the ability to convey classification rationale deteriorates
Solution Approach 1:
The patent segments the image into tiles and provides classification rationale at the tile level rather than requiring interpretation of every individual pixel. Each tile's classification is based on its local features, making the rationale more interpretable while collectively covering the entire image through aggregation of tile-level explanations.
Solution Approach 2:
The patent introduces tile-level classifications as an intermediary between pixel-level processing and whole-slide interpretation. The Grad-CAM visualizations are generated at the tile level, serving as interpretable intermediaries that convey the rationale for each local classification without requiring direct interpretation of all pixels, thus preserving classification rationale while maintaining complete coverage.
Data Source
AI summary
Accurate tissue segmentation is performed without a priori knowledge of tissue type or other extrinsic information not found within the subject image, and may be combined with classification analysis so that diseased tissue is not only delineated within an image but also characterized in terms of disease type. In various embodiments, a source image is decomposed into smaller overlapping subimages such as square or rectangular tiles, which are sifted based on a visual criterion. The visual criterion may be one or more of image entropy, density, background percentage, or other discriminator. A convolutional neural network produces tile-level classifications that are aggregated to produce a tissue segmentation and, in some embodiments, to classify the source image or a subregion thereof.