Hierarchical Medical Image Annotation System for Learning Data
Solution Overview
Problem
Current methods for generating high-quality annotation data for medical images require significant manual labor, making it inefficient to create learning data that reflects the hierarchy and relationship of annotation information necessary for accurate automatic diagnosis support.
Innovation Solution
An annotation support apparatus and method that utilizes a hierarchical structure to manage medical images, allowing for the efficient input, storage, and association of annotation information, including patient and finding information, to facilitate the generation of learning data for machine learning models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If manual annotation input is used for medical images, then annotation information can be acquired, but the labor required is large and the process is inefficient
Solution Approach 1:
The system performs preliminary actions by automatically generating candidate annotation information (such as lesion locations, types, and characteristics) using image analysis algorithms before the user provides final confirmation. This pre-processing reduces the manual workload significantly while maintaining annotation quality.
Solution Approach 2:
The system enables self-service annotation by allowing the automated annotation generation process to handle the majority of annotation tasks without continuous manual intervention. Users only need to review and confirm the generated annotations, making the system serve itself for the complex analysis portions while human oversight ensures accuracy.
2Reliability
If annotation information is stored without hierarchical structure, then storage is simple, but the hierarchy and relationship of annotation information cannot be reflected
Solution Approach 1:
The annotation information is segmented into multiple hierarchical levels: patient-level information (demographics, medical history), examination-level information (examination type, date), image-level information (image quality, modality), and lesion-level information (location, type, characteristics). This segmentation allows the system to manage complex relationships while maintaining organized data structures that reflect the natural hierarchy of medical diagnostic information.
Solution Approach 2:
The system adds hierarchical dimensions to the data structure by organizing annotation information across multiple levels (patient → examination → image → lesion). This dimensional organization transforms flat storage into a multi-layered structure that captures the relationships between different types of annotation information, enabling more accurate diagnosis support models.
3Manufacturing precision
If comprehensive annotation information is collected, then learning data quality improves, but the complexity of data processing increases
Solution Approach 1:
Comprehensive annotation information is segmented into distinct categories (patient information, examination information, image information, lesion information) with clearly defined data schemas for each category. This segmentation allows the system to collect detailed information while managing processing complexity through structured data organization and category-specific processing pipelines.
Data Source
AI summary
An annotation support apparatus manages a medical image on the basis of an attribute of each medical image, using a hierarchical structure that has one medical image as a node and includes a medical image hierarchy to which a node of the medical image belongs and at least one of an upper hierarchy higher than the medical image hierarchy or a lower hierarchy lower than the medical image hierarchy. The annotation support apparatus includes: an image acquisition unit that acquires the medical image; a display unit; an annotation input receiving unit that receives designation of any one of the node of the medical image or a node associated with the node of the medical image in the hierarchical structure and an input of annotation information related to the designated node; and an annotation storage unit that stores the annotation information to be associated with the designated node.