Voxel-Level Patch Classification for Medical Image Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Manual assessment of high-resolution medical images, such as 3D and 4D CT scans, is clinically infeasible due to time constraints, necessitating automated and semi-automated processing methods for tasks like nodule detection and segmentation, which often require extensive training data and complex feature sets when using multi-scale machine learning approaches.
Innovation Solution
A voxel-level machine-learnt classifier is employed, where a processor segments anatomical structures using non-ML methods for large structures and applies ML to small patches, allowing for cloud-based classification, feedback, and the creation of a labeled library for faster throughput and reduced complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multi-scale machine learning approaches are used for segmentation, then segmentation accuracy is improved, but processing complexity and training data requirements increase
Solution Approach 1:
The patent divides the image processing task into multiple scales: a first scale for locating gross anatomical structures using non-ML methods, and a second (voxel) scale for detailed segmentation using ML on small patches. This hierarchical segmentation resolves the contradiction by applying complexity only where needed rather than across the entire image.
Solution Approach 2:
The patent applies machine learning methods locally to small patches at the voxel level rather than globally across the entire image. This localized application maintains high segmentation accuracy for detailed structures while reducing overall processing complexity and training data requirements compared to applying ML uniformly at all scales.
2Measurement precision
If multi-scale machine learning approaches are used for segmentation, then segmentation accuracy is improved, but the amount of training data and feature sets required increases
Solution Approach 1:
By segmenting the processing task into coarse localization (non-ML) and fine segmentation (ML on patches), the patent reduces the total volume of training data needed compared to applying multi-scale ML uniformly, since the ML component only processes small patches rather than entire images at multiple scales.
Solution Approach 2:
The patent applies ML locally to small patches rather than globally, which reduces the quantity of training data required while maintaining segmentation accuracy for the detailed voxel-level structures that require precise classification.
3Measurement precision
If manual assessment of medical images is performed, then diagnostic accuracy is maintained, but time consumption increases making it clinically infeasible
Solution Approach 1:
The patent segments the processing into automated coarse localization and automated fine segmentation with minimal user interaction, eliminating manual assessment entirely while maintaining diagnostic accuracy through the hierarchical approach that combines robust non-ML localization with ML-based detailed segmentation.
Solution Approach 2:
The system performs both localization and detailed segmentation automatically without requiring manual intervention, making the process clinically feasible by eliminating time-consuming manual assessment while maintaining diagnostic accuracy through the combined non-ML and ML approach.
4Productivity
If cloud-based classification with small patches is used, then processing speed and scalability are improved, but data transmission requirements increase
Solution Approach 1:
The patent extracts only the necessary small patches from the larger medical images and transmits only these patches to the cloud for classification, rather than transmitting entire images. This reduces data transmission volume while maintaining processing speed and scalability benefits of cloud-based computation.
Solution Approach 2:
By processing only small local patches in the cloud rather than entire images, the patent reduces data transmission requirements while maintaining high processing speed and scalability, as the cloud only needs to handle small, manageable data units rather than large complete images.
Data Source
AI summary
A single level machine-learnt classifier is used in medical imaging. A gross or large structure is located using any approach, including non-ML approaches such as region growing or level-sets. Smaller portions of the structure are located using ML applied to relatively small patches (small relative to the organ or overall structure of interest). The classification of small patches allows for a simple ML approach specific to a single scale or at a voxel/pixel level. The use of small patches may allow for providing classification as a service (e.g., cloud-based classification) since partial image data is to be transmitted. The use of small patches may allow for feedback on classification and updates to the ML. The use of small patches may allow for the creation of a labeled library of classification partially based on ML. Given a near complete labeled library, a simple matching of patches or a lookup can replace ML classification for faster throughput.