Endoscopic Lesion Detection Training for Region-Aware Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial neural network models for detecting lesions in endoscopic images face challenges due to noise, diverse lesion shapes, and varying image environments, leading to low sensitivity and accuracy.
Innovation Solution
Training data is generated considering the characteristics of endoscopic image regions and lesions, including clinical characteristics and using a loss function with a DIoU structure, and the model is trained to detect both lesions and non-lesion images, with weight assignment based on reading difficulty.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If existing artificial neural network models are used for lesion detection, then the detection process can be automated, but the sensitivity and accuracy remain low due to noise, diverse lesion shapes, and varying image environments
Solution Approach 1:
The patent applies local quality by creating region-specific training data that accounts for different characteristics of endoscopic image regions. The training process considers local variations in image quality, noise patterns, and lesion appearances across different anatomical regions, enabling the model to adapt to local conditions rather than applying a uniform detection approach throughout the entire image
Solution Approach 2:
The patent employs parameter changes by modifying the training data to reflect real-world variations in lesion characteristics, image quality, and environmental conditions. The loss function with DIoU structure and weighted training samples adjust the optimization parameters to better match clinical reality, improving the model's sensitivity and accuracy in detecting diverse lesion types
2Measurement precision
If training data is generated considering clinical characteristics and region characteristics, then sensitivity and accuracy improve, but data preparation complexity increases
Solution Approach 1:
The patent applies preliminary action by pre-processing and organizing training data to reflect clinical characteristics before model training. Region characteristics and lesion features are pre-analyzed and encoded into the training dataset structure, including the loss function design with DIoU and weighted samples, so that the model receives optimized training inputs without requiring complex real-time adjustments during deployment
3Reliability
If the model is trained to detect both lesions and non-lesion images with weight assignment based on reading difficulty, then detection performance improves, but training time and computational resources increase
Solution Approach 1:
The patent applies partial or excessive action by implementing weighted training samples that focus computational effort on more challenging cases. The loss function assigns higher weights to difficult-to-detect lesions and regions with higher reading difficulty, allowing the model to prioritize learning from critical examples while still processing the full dataset, thereby improving reliability without requiring excessive training iterations
Data Source
Figure 1~2
Figure 3
Figure 4(a)~4(d)
AI summary
The present disclosure is directed to a method of training an artificial neural network model for detecting lesions in endoscopic images that is performed by a computing device including at least one processor. The method includes: generating training data including labels for lesions based on endoscopic images with characteristics of regions where the endoscopic images are captured and characteristics of the lesions taken into consideration; and training an artificial neural network model to detect the lesions in endoscopic images based on the training data.