Tile-Based AI Image Recognition for Pathology
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current AI integration in pathological medicine systems faces challenges due to non-standardized software systems and varying DICOM image resolutions, leading to compatibility issues and high resource requirements for high-resolution image inference, making it impractical for busy pathologists.
Innovation Solution
The method involves cutting images into uniform-sized tiles for input into a trained AI model, reducing hardware burden and increasing inference speed, and includes a system for regular or event-triggered image recognition with structured report output and heat map display.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-resolution DICOM images are input into AI models, then image recognition accuracy is improved, but hardware resource requirements and inference time increase significantly
Solution Approach 1:
The patent divides high-resolution DICOM images into multiple smaller tiles with uniform sizes before inputting them into the AI model. This segmentation approach allows the model to process images in manageable chunks, reducing the hardware resource requirements and inference time for each individual tile while maintaining the ability to accurately analyze the complete image through aggregation of tile results.
2Measurement precision
If high-resolution DICOM images are input into AI models, then image recognition accuracy is improved, but inference time increases to more than 10 minutes per image
Solution Approach 1:
The patent segments high-resolution images into multiple smaller tiles, which significantly reduces the inference time for each tile. By processing tiles in parallel and aggregating their results, the system achieves both high accuracy and reduced total inference time compared to processing entire high-resolution images sequentially.
Solution Approach 2:
The patent performs preliminary actions by converting image formats to specified formats and cutting images into uniform-sized tiles before they are input into the AI model. These preparatory steps optimize the image data structure, enabling faster and more efficient inference while maintaining recognition accuracy.
3Productivity
If images are cut into multiple uniform-sized tiles, then hardware resource burden is reduced and inference speed increases, but image recognition accuracy may be compromised
Solution Approach 1:
The patent merges the results from multiple individual tile predictions by aggregating their judgment results to produce the final image-level recognition output. This combining process ensures that the overall accuracy is maintained by synthesizing information from all tiles, while each tile benefits from the computational efficiency of processing smaller images.
Solution Approach 2:
The patent applies segmentation by cutting images into uniform-sized tiles that are then processed independently through the AI model. This approach enables parallel processing and reduces the computational burden on hardware resources while maintaining accuracy through systematic aggregation of tile-level predictions.
4Adaptability or versatility
If AI models process DICOM images with varying resolutions, then compatibility with different scanners is achieved, but model compatibility issues arise
Solution Approach 1:
The patent segments images from different scanners into uniform-sized tiles, which standardizes the input format for the AI model. This segmentation approach allows the model to receive consistent input dimensions regardless of the original image resolution or scanner type, thereby improving compatibility while reducing the complexity of handling varying resolutions.
Solution Approach 2:
The patent changes the parameter of image resolution by converting and resizing images to uniform sizes before processing. This parameter transformation enables the AI model to handle images from different scanners consistently, improving adaptability while simplifying the model's requirements and reducing compatibility issues.
Data Source
AI summary
A method and a system for image recognition and a computer readable storage medium are provided. The method includes: obtaining an image to be recognized; cutting the image to be recognized into multiple tiles, wherein a size of each of the tiles conforms to a preset size; inputting the tiles into a trained artificial intelligence model respectively; obtaining multiple judgment results of the tiles, wherein each of the judgment results includes a specified category in which the corresponding tile is classified into multiple categories; and storing the judgment results as a structured report to be output to a digital pathology platform.