Multiscale Image Annotation via Selective User Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Manual annotation of pathology images for training machine learning algorithms is time-consuming, expensive, and prone to errors, requiring substantial user effort to correct erroneous segmentation results.
Innovation Solution
A system and method that utilize a multiscale image representation, allowing users to provide selective feedback through a configurable viewing window with a magnification factor and spatial offset parameter, guiding the user to areas where their input significantly changes the learned annotation by the machine learning algorithm, thereby optimizing the training process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual annotation is performed on pathology images to train machine learning algorithms, then the training data quality is improved, but the time consumption and cost increase significantly
Solution Approach 1:
The system implements an interactive feedback mechanism where users correct segmentation errors and the system learns from these corrections. The machine learning algorithm processes user feedback and updates its annotations, creating a feedback loop that improves training data quality while reducing the burden on users to manually annotate entire images from scratch.
Solution Approach 2:
The system performs self-service by automatically generating initial annotations and allowing users to only correct errors rather than creating annotations from scratch. The machine learning algorithm autonomously processes and learns from user feedback, reducing the overall effort required from users while maintaining high training data quality.
2Reliability
If manual annotation is performed by pathologists, then the annotation reliability is improved, but the cost increases significantly
Solution Approach 1:
The system uses feedback from pathologists to correct segmentation errors, where the machine learning algorithm learns from these corrections. This feedback mechanism maintains high annotation reliability by incorporating expert input while reducing the quantity of manual annotation work required, thereby lowering costs.
Solution Approach 2:
The machine learning algorithm acts as an intermediary between pathologists and the final annotation output. It processes expert feedback and generates annotations, reducing the direct workload on pathologists while maintaining the reliability of annotations that would otherwise require extensive manual work.
3Measurement precision
If users correct all erroneous segmentation results, then the annotation accuracy is improved, but the effort required increases substantially
Solution Approach 1:
The system implements a feedback mechanism where users only need to correct erroneous segmentation results rather than review and annotate entire images. The machine learning algorithm processes this selective feedback and updates its annotations, achieving high annotation accuracy with substantially reduced user effort.
Solution Approach 2:
Instead of requiring users to perform complete manual annotation of entire images, the system applies partial action by only requiring corrections where the machine learning algorithm makes errors. This partial correction approach achieves sufficient annotation accuracy without the excessive effort of full manual annotation.
Data Source
AI summary
A system and method are provided which use a machine learning algorithm to obtain a learned annotation of objects in one or more scales of a multiscale image. A viewing window (300) is provided for viewing the multiscale image. The viewing window is configurable on the basis of a magnification factor, which selects one of the plurality of scales for viewing, and a spatial offset parameter. A user may provide a manual annotation of an object in the viewing window, which is then used as training feedback in the learning of the machine learning algorithm. To enable the user to more effectively provide the manual annotation, the magnification factor and the spatial offset parameter for the viewing window may be automatically determined, namely by the system and method determining where in the multiscale image the manual annotation of the object would have sufficient influence on the learned annotation provided by the machine learning algorithm. The determined influence may be shown in the form of an overlay (350) in the viewing window.