Endoscope Image Learning Using Biopsy-Based Ground Truth
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The accuracy of ground-truth data for medical images varies among users, leading to inconsistent training of machine learning recognizers for medical image analysis, making it difficult to effectively train these recognizers for tasks like lesion detection and malignancy identification.
Innovation Solution
A learning apparatus and method that utilizes biopsy information to generate ground-truth region data by setting regions where biopsies have been performed as ground-truth regions, enabling accurate training of neural network-based recognizers for medical image analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If ground-truth data is generated by multiple professional users (doctors) independently, then the data collection process is efficient and scalable, but the accuracy and consistency of the ground-truth data decreases due to varying definitions and interpretations
Solution Approach 1:
The patent segments the ground-truth data generation process into two distinct phases: (1) efficient parallel annotation by multiple doctors to generate candidate ground-truth data, and (2) subsequent consistency resolution through comparison and selection. This segmentation allows maintaining high productivity during data collection while ensuring accuracy through structured validation
Solution Approach 2:
The patent implements a feedback mechanism where the recognizer's detection results are compared against the ground-truth data generated by multiple doctors. The system uses this feedback to identify inconsistencies and select or refine the most accurate ground-truth annotations, thereby improving data accuracy while maintaining collection efficiency
2Measurement precision
If ground-truth data is generated by a single user, then the accuracy and consistency of the data improves, but the productivity and scalability of data collection decreases
Solution Approach 1:
The patent merges the strengths of multiple annotators by combining their annotations into a unified ground-truth dataset. Through comparison and consensus-building mechanisms, the system integrates multiple perspectives while maintaining high accuracy, achieving both productivity and precision that would be difficult to obtain from a single annotator
3Duration of action of moving object
If the recognizer is trained on ground-truth data with varying definitions from different users, then the training process is faster with more diverse data, but the learning effectiveness and recognizer performance deteriorates
Solution Approach 1:
The patent performs preliminary resolution of ground-truth inconsistencies before the training process begins. By establishing consistent and accurate ground-truth data through comparison and selection of annotations from multiple doctors prior to training, the system ensures that the recognizer receives high-quality training data, thereby maintaining both training efficiency and learning effectiveness
Data Source
AI summary
Provided are a learning apparatus, a learning method, a program, a trained model, and an endoscope system that perform effective learning by using accurate ground-truth data. The learning apparatus 14 is a learning apparatus 14 having a recognizer implemented by a neural network, and a processor. The processor (CPU 41) acquires a learning image 53 and biopsy information associated with the learning image 53, the learning image 53 being obtained by capturing an image of an examination object, the biopsy information being information indicating a location where a biopsy for the examination object has been performed, generates ground-truth region data in which a region including the location where the biopsy is performed is set as a ground-truth region, on the basis of the biopsy information, and trains the recognizer that recognizes a region of interest by using the learning image 53 and the ground-truth region data.