Lesion Boundary Refinement Using Uncertainty Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine-learning algorithms for medical image segmentation face inaccuracies in lesion boundary identification due to incomplete or incorrect annotations, which are exacerbated by the need to upsample lower resolution images and differences in training data and loss functions, leading to discrepancies in boundary identification among algorithms.
Innovation Solution
A method that utilizes uncertainty data to identify and correct lesion boundaries in medical image data by generating probability and uncertainty data points, allowing for the use of region growing and shrinking algorithms to refine lesion boundaries, and optionally involving user input for validation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine-learning algorithms are trained on annotated medical image data, then lesion identification capability is improved, but boundary accuracy deteriorates due to incomplete or incorrect annotations
Solution Approach 1:
The system generates uncertainty maps that provide feedback on the reliability of predicted boundaries. High uncertainty regions indicate where the model is less confident, allowing for iterative refinement and correction of boundary annotations, thereby improving boundary accuracy while maintaining lesion identification capability
Solution Approach 2:
Uncertainty maps serve as an intermediary between the machine-learning algorithm and the final boundary identification. These maps quantify the model's confidence and guide further annotation or correction, resolving the contradiction by providing a measure of reliability without requiring perfect initial annotations
2Stability of the object's composition
If lower resolution images are upscaled to uniform resolution for training, then training data consistency is improved, but boundary precision deteriorates
Solution Approach 1:
The system applies different processing qualities to different regions: high-resolution processing is applied specifically to boundary regions where precision is critical, while lower-resolution processing is acceptable for interior regions. This local differentiation maintains training consistency while preserving boundary precision
3Adaptability or versatility
If multiple machine-learning algorithms are used to identify boundaries, then comprehensive coverage is improved, but agreement on precise boundaries deteriorates
Solution Approach 1:
The system combines multiple algorithm outputs by generating uncertainty maps that aggregate confidence information across different algorithms. Regions with high uncertainty indicate disagreement between algorithms, allowing for targeted resolution while maintaining the benefits of multiple perspectives
Solution Approach 2:
The uncertainty maps provide feedback on inter-algorithm agreement, highlighting regions where boundaries differ significantly. This feedback enables iterative refinement and consensus-building, improving boundary agreement while preserving the comprehensive coverage provided by multiple algorithms
Data Source
AI summary
The present invention provides a method, computer program and processing system for identifies boundaries of lesions within image data. The image data is processed using a machine learning algorithm to generate probability data and uncertainty data. The probability data provides, for each image data point of the image data, a probability data points indicating a probability that said image data point is part of a lesion. The uncertainty data provides, for each probability data point, an uncertainty data point indicating an uncertainty of the said probability data point. The uncertainty data is processed to identify or correct boundaries of the lesions.