Ultrasound Needle Segmentation via Hard-to-Soft Label Transformation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Ultrasound imaging is hindered by needle reverberation artifacts, which are challenging to identify and distinguish from actual tissue due to their ambiguous boundaries and varying intensity distributions, leading to confusion in medical image analysis and annotation disagreements among experts.
Innovation Solution
A method involving machine-learning models, including a first model for hard-labeling and a second U-Net architecture for soft-labeling, to segment images of needles by generating mean and standard deviation maps, clustering artifacts, and adjusting pixel values based on exponential decay, effectively distinguishing needle artifacts from tissue.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If pixel-wise labeling is performed to identify needle reverberation artifacts, then annotation accuracy can be improved, but the time and effort required for annotation increases significantly
Solution Approach 1:
The patent applies preliminary action by using automated algorithms to pre-identify and label needle reverberation artifacts before human annotators review the images. This pre-labeling process provides a head start on the annotation task, reducing the time required while maintaining accuracy through human verification of the pre-generated labels.
Solution Approach 2:
The patent introduces an intermediary automated detection system that acts as a mediator between the raw ultrasound images and the final annotations. This intermediary system generates preliminary artifact labels that guide human annotators, reducing their cognitive load and annotation time while preserving accuracy through the collaborative human-machine process.
2Reliability
If multiple annotators label needle reverberation artifacts independently, then annotation reliability can be improved through consensus, but disagreement and inconsistency increase due to ambiguous artifact boundaries
Solution Approach 1:
The patent implements feedback mechanisms where annotators' labels are compared against each other and against automated detection results. Discrepancies trigger review and reconciliation processes, creating a feedback loop that progressively improves annotation consistency. The system provides feedback on inter-annotator agreement metrics to guide the annotation process.
Solution Approach 2:
The patent segments the annotation process into distinct stages: automated pre-detection, initial human labeling, consensus building through comparison, and final verification. This segmentation of the annotation workflow allows each stage to focus on specific tasks, improving overall reliability while managing complexity through structured process division.
3Ease of manufacture
If traditional image processing methods are used to remove needle reverberation artifacts, then processing simplicity is maintained, but artifact removal accuracy decreases due to ambiguous boundaries and varying intensity distributions
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting processing thresholds and parameters based on local image characteristics such as intensity distribution, texture patterns, and spatial context. Rather than using fixed thresholds, the system adapts parameters to match the varying properties of different artifact regions, improving removal accuracy while maintaining reasonable processing simplicity through automated parameter selection.
Solution Approach 2:
The patent implements local quality by applying different processing strategies and parameters to different regions of the image based on their specific characteristics. Needle reverberation artifacts in different locations have different intensity distributions and boundary characteristics, so the system tailors the artifact removal approach to each local region, improving overall accuracy while keeping the overall process manageable through regional specialization.
Data Source
AI summary
Provided are systems, methods, and computer program products for segmenting an image. A method includes segmenting each image in a sequence of images including a needle into a needle and at least one needle artifact based on processing each image with a first machine-learning model trained with a plurality of hard labels for a plurality of images, resulting in a plurality of hard-labeled images, transforming each hard-labeled image of the plurality of hard-labeled images into a soft-labeled image including pixel values corresponding to an effect of the at least one needle artifact, resulting in a plurality of soft-labeled images, and segmenting at least one image of the sequence of images based on processing the at least one image with a second machine-learning model trained at least partially with the plurality of soft-labeled images.