Lesion Proposal Generator Conditioning for Medical Image Annotation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing medical image databases, such as DeepLesion, suffer from uncertainties and incomplete annotations, making them ill-suited for training machine learning systems, as they often lack comprehensive lesion annotations, which are costly and time-consuming to create.
Innovation Solution
A method and system that condition a lesion proposal generator (LPG) using a subset of annotated images to enhance lesion annotations, integrating it with a selective lesion proposal classifier (LPC) to reduce false positives and harvest additional annotations from unannotated images, forming 3D bounding boxes for improved lesion detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If comprehensive manual annotation is performed on all medical images, then annotation completeness improves, but time consumption and cost increase significantly
Solution Approach 1:
The system performs preliminary automated annotation using the conditioned LPG to generate initial lesion proposals before any manual review. This preliminary action creates a draft annotation set that covers most lesions, reducing the subsequent manual annotation workload while maintaining high completeness.
Solution Approach 2:
The system uses harvested annotations from automatically annotated images to iteratively improve and re-condition the LPG, enabling it to perform better annotation tasks autonomously. This self-improving mechanism reduces dependency on continuous manual annotation while progressively enhancing annotation quality and completeness.
2Productivity
If automated lesion detection is applied to all images, then annotation speed improves, but false-positive rate increases
Solution Approach 1:
The system implements feedback loops where harvested annotations from automated detection are used to re-condition the LPG, improving its accuracy. Additionally, the system iteratively refines detection parameters based on performance metrics, continuously reducing false positives while maintaining high detection speed across the dataset.
Solution Approach 2:
The system transitions from 2D slice-based annotation to 3D volumetric annotation, leveraging spatial relationships across multiple slices to distinguish true lesions from false positives. This dimensional enhancement provides contextual information that reduces false-positive rates while maintaining automated processing speed.
3Ease of manufacture
If 2D slice-based annotation is used, then annotation process is simple, but lesion detection recall is insufficient
Solution Approach 1:
The system extends annotation from 2D slices to 3D volumes by aggregating lesion proposals across multiple slices and applying spatial consistency constraints. This 3D approach captures lesions that span multiple slices, significantly improving detection recall while building upon the simpler 2D annotation foundation through iterative refinement.
4Quantity of substance
If existing medical image databases are used for training, then data availability improves, but annotation quality deteriorates due to uncertainties and incomplete annotations
Solution Approach 1:
The system extracts and isolates high-confidence lesion annotations from the conditioned LPG output, separating them from low-confidence or false-positive proposals. This extraction process creates a curated subset of reliable annotations from the larger automated output, improving overall annotation quality while utilizing the abundant data from existing databases.
Data Source
AI summary
A method of harvesting lesion annotations includes conditioning a lesion proposal generator (LPG) based on a first two-dimensional (2D) image set to obtain a conditioned LPG, including adding lesion annotations to the first 2D image set to obtain a revised first 2D image set, forming a three-dimensional (3D) composite image according to the revised first 2D image set, reducing false-positive lesion annotations from the revised first 2D image set according to the 3D composite image to obtain a second-revised first 2D image set, and feeding the second-revised first 2D image set to the LPG to obtain the conditioned LPG, and applying the conditioned LPG to a second 2D image set different than the first 2D image set to harvest lesion annotations.