Medical Image Learning Method for Region of Interest Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing medical image learning methods face inefficiencies and inaccuracies due to the collection of large numbers of similar and non-relevant images, leading to class imbalance issues and increased storage demands.
Innovation Solution
A medical image learning method that involves performing abnormality detection using a first model trained on normal images, sorting extracted images to prevent erroneous recognition, and generating a second model through second learning using a sorted image group to detect regions of interest with high accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a large number of frame images are collected from motion picture recording for learning data, then the coverage of various variations is improved, but the storage load increases and class imbalance problem occurs
Solution Approach 1:
The patent extracts only the necessary frame images containing regions of interest from the motion picture recording, rather than using all frame images. This extraction process removes unnecessary data (frames without lesions) while preserving useful information (frames with lesions), thereby reducing storage requirements while maintaining learning effectiveness.
Solution Approach 2:
The patent performs preliminary sorting and selection of frame images before using them for learning. By pre-identifying and selecting only the necessary frames that contain regions of interest, the system prepares the learning data in advance, avoiding the need to store and process all frame images from the motion picture recording.
2Quantity of substance
If similar images are excessively collected from motion picture frames, then the quantity of learning data increases, but the learning efficiency decreases due to redundant information
Solution Approach 1:
The patent extracts only unique and necessary frame images that contain regions of interest, removing redundant similar images from the learning dataset. This extraction ensures that each frame image contributes meaningful information to the learning process without duplicating existing data.
Solution Approach 2:
The patent changes the selection parameters for frame image collection by using sorting based on similarity metrics and region of interest detection. Instead of collecting all frames or using simple random sampling, the system applies parameter-based selection to identify and collect only the most valuable frames for learning.
3Quantity of substance
If random reduction of collected images is performed to avoid storage pressure, then the storage load is reduced, but the accuracy may decrease due to loss of useful image data
Solution Approach 1:
The patent extracts frame images based on specific criteria (presence of regions of interest, similarity sorting) rather than random selection. This targeted extraction ensures that useful image data is preserved while reducing storage requirements, avoiding the accuracy degradation that would result from random reduction.
Solution Approach 2:
The patent performs preliminary analysis and sorting of frame images to identify and preserve those containing regions of interest before reduction. By pre-processing the data to identify valuable frames, the system ensures that accuracy-critical images are retained in the reduced dataset.
4Stability of the object's composition
If all frame images including those without regions of interest are used for learning, then the class balance improves, but the learning efficiency decreases due to inclusion of non-useful images
Solution Approach 1:
The patent extracts and removes frame images without regions of interest from the learning dataset. By eliminating these non-useful frames, the system improves learning efficiency while maintaining an appropriate balance between classes through selective extraction of frames containing lesions.
Data Source
AI summary
A trained first model is generated through first learning using a first learning image group constituted of a normal image which is a medical image having no region of interest. An input image group including at least the medical image different from the first learning image group is input to the trained first model, and abnormality detection is performed. The extracted image used for learning to prevent erroneous recognition is sorted according to a result of the abnormality detection, and second learning using a second learning image group including at least the extracted image is performed. A second model that detects the region of interest is generated through the second learning.