Medical Image ROI Extraction Using Dynamic Patch Sizing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image processing techniques for medical data face challenges in accurately and efficiently extracting regions of interest of varying sizes, as they often require fixed image sizes, leading to decreased accuracy and increased computational time due to the need to process large numbers of subregions.
Innovation Solution
An image processing apparatus that uses a processing circuitry to extract regions of interest by setting a specified point in medical image data, employing a 3D U-Net inference model to generate probability maps, and adjusting binary thresholds to ensure accurate extraction of regions of interest regardless of size, while minimizing computational time by expanding processing ranges as needed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If fixed-size image processing is used with deep neural networks, then extraction accuracy is improved for small regions of interest, but extraction accuracy deteriorates for large regions of interest
Solution Approach 1:
The patent divides the medical image into multiple patches of different sizes based on the estimated size of the region of interest. Small ROIs are processed using small patches while large ROIs are processed using large patches, allowing the system to adapt the processing scale to the actual ROI size and maintain high extraction accuracy across different scales.
Solution Approach 2:
The patent dynamically adjusts the patch size according to the estimated ROI size. Instead of using a fixed patch size, the system selects from multiple patch sizes (e.g., 64×64, 128×128, 256×256) based on the predicted ROI dimensions, making the processing adaptable to varying ROI scales while maintaining accuracy.
2Manufacturing precision
If the fixed processing size is made larger to accommodate large regions of interest, then extraction completeness is improved, but extraction accuracy deteriorates for small regions of interest
Solution Approach 1:
The patent segments the image into multiple patches of different sizes rather than using a single large fixed-size patch. This allows small ROIs to be processed with small patches (maintaining accuracy) while large ROIs are processed with appropriately sized larger patches (ensuring completeness), resolving the trade-off between accuracy and completeness.
Solution Approach 2:
The patent applies different processing configurations to different regions of the image based on local ROI characteristics. Each patch is processed with a size appropriate to the local ROI size, rather than applying a uniform global processing configuration, thereby optimizing both accuracy and completeness locally.
3Manufacturing precision
If the entire image is divided into many fixed-size subregions for processing, then complete coverage is improved, but computational time increases significantly
Solution Approach 1:
The patent performs preliminary estimation of the ROI size before dividing the image into patches. Based on this preliminary information, the system determines the appropriate patch size and number of patches needed, avoiding unnecessary division into excessive small patches and thereby reducing computational time while ensuring complete coverage.
Solution Approach 2:
The patent uses partial action by processing only the necessary portions of the image with appropriately sized patches rather than uniformly processing the entire image with fine-grained small patches. This selective processing approach reduces the total number of patches to process while ensuring complete ROI coverage.
4Measurement precision
If multiple inference models with different fixed sizes are used to handle various ROI sizes, then extraction accuracy across different sizes is improved, but device complexity increases
Solution Approach 1:
The patent uses a single inference model that can process multiple patch sizes rather than requiring separate inference models for each size. The model is designed to accept variable input sizes, making it universal and multi-functional, thereby maintaining extraction accuracy across different ROI sizes while avoiding the complexity of managing multiple specialized models.
Solution Approach 2:
The patent employs a single flexible inference model that dynamically adapts to different input patch sizes rather than requiring static specialized models for each size. This dynamic capability allows the system to maintain accuracy across varying ROI sizes while simplifying the overall system architecture.
Data Source
AI summary
An image processing apparatus according to embodiments extracts a region of interest including a specified point on medical image data. The image processing apparatus includes processing circuitry. The processing circuitry extracts a first extraction region estimated as the region of interest from first partial image data included in a first processing range that includes the specified point of the medical image data. The processing circuitry extracts a second extraction region estimated as the region of interest from second partial image data included in a second processing range that is in contact with an edge portion of the first processing range or that includes at least a part of the edge portion, from the medical image data when the first extraction region is determined as a part of the region of interest.