One-Shot Medical Image Feature Localization Using Spatial Templates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current techniques for generating annotated training data for medical image inferencing models are inefficient, burdensome, and prone to error, particularly due to the high cost and time required for manual annotation, and existing automated methods struggle to accurately localize features with similar visual characteristics in medical images.
Innovation Solution
Combining pretrained foundation models with domain knowledge to leverage reference spatial relationships between anatomical features, enabling one-shot localization and labeling of target features in medical images by using a single labeled template image, and restricting the search region based on known spatial positions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual annotation is used to generate annotated training data, then annotation accuracy is improved, but time consumption and cost increase significantly
Solution Approach 1:
The patent introduces a pre-trained foundation model as an intermediary that performs initial feature extraction and localization on medical images. This intermediary model processes the images to generate candidate feature locations, which are then refined through matching with reference pixel features from template images. This intermediary processing step significantly reduces the time and cost of manual annotation while maintaining acceptable accuracy through subsequent automated refinement steps.
Solution Approach 2:
The patent applies preliminary action by using pre-trained foundation models that have already been trained on large datasets before being applied to the specific medical image annotation task. These pre-trained models perform initial feature extraction and provide starting points for localization, eliminating the need to train models from scratch for each annotation task. This preliminary preparation significantly accelerates the annotation process while maintaining quality.
2Measurement precision
If manual annotation is used to generate annotated training data, then annotation quality is improved, but financial cost increases significantly
Solution Approach 1:
The patent employs copying by using reference pixel features extracted from template images (which may be manually annotated once) to generate annotations for multiple target images. Instead of manually annotating each image individually, the system copies and adapts feature patterns from reference templates to new images through pixel feature matching. This copying approach dramatically reduces financial costs while maintaining annotation quality through the matching process.
Solution Approach 2:
The patent implements self-service by enabling the annotation system to automatically generate annotations for medical images using pre-trained foundation models and reference templates without requiring continuous human intervention. The system serves itself by autonomously performing feature extraction, matching, and localization across multiple images, reducing the need for expensive manual annotation services while maintaining quality through automated refinement.
3Productivity
If automated annotation methods are used, then processing efficiency is improved, but accuracy in localizing features with similar visual characteristics deteriorates
Solution Approach 1:
The patent applies local quality by focusing the matching process on specific local regions around candidate feature locations rather than searching the entire image. Once the foundation model identifies candidate features, the system extracts pixel features from localized regions around these candidates and compares them with reference pixel features from template images. This localized approach improves accuracy for features with similar visual characteristics by concentrating computational effort on relevant local patterns rather than performing global searches.
Solution Approach 2:
The patent employs segmentation by dividing the annotation task into distinct stages: (1) candidate feature identification using pre-trained foundation models, (2) pixel feature extraction from localized regions around candidates, (3) matching with reference pixel features from templates, and (4) final localization refinement. This segmentation allows each stage to specialize in specific aspects of the problem, improving overall accuracy for difficult features while maintaining high processing efficiency through automated multi-stage processing.
Data Source
AI summary
One-shot medical image feature localization techniques are provided that employ pretrained foundation models and domain knowledge. In an example, a computer-implemented method can comprise determining positions of target features within a target medical image of an anatomical region of a subject based on reference spatial relationships between the target features as defined in reference spatial information, and based on matching reference pixel features respectively associated with the target features with corresponding subsets of pixel features of the target medical image, wherein the reference pixel features comprise template image pixel features extracted from labeled versions of the target features as included in a template medical image depicting the anatomical region of a reference subject. The method further comprises generating label information for the target features identifying the target features and their positions and associating the label information with the target medical image.