Diagnosis Support Device Lesion Detection via Image Retrieval
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting lesion candidate regions in medical images using machine learning require large datasets and struggle to detect various lesion types in a general-purpose manner.
Innovation Solution
A diagnosis support device that receives input retrieval words, retrieves medical images associated with key findings, classifies images based on similarity, and detects lesion candidate regions in examination images using a combination of similarity classification and optional trained models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning models are used to detect lesion candidate regions, then detection accuracy for specific lesions is improved, but the system cannot detect various lesion types in a general-purpose manner and requires large datasets
Solution Approach 1:
The patent creates a universal lesion detection system by retrieving and classifying multiple types of medical images (CT, MRI, ultrasonic images) along with their associated key findings and diagnosis names. The system builds a comprehensive database that can detect various lesion types (tumors, cysts, abscesses, etc.) across different organs and image modalities, enabling general-purpose detection without requiring separate specialized models for each lesion type.
Solution Approach 2:
The patent introduces an intermediary classification mechanism that groups retrieved medical images by similarity and associates them with key findings and diagnosis names. This intermediary layer bridges the gap between raw image data and lesion detection, allowing the system to infer lesion characteristics from similar cases without direct machine learning training, thus achieving versatility without requiring large specialized datasets.
2Reliability
If machine learning models are trained on large datasets, then detection performance is improved, but the system becomes complex and difficult to implement for various lesion types
Solution Approach 1:
Instead of training complex machine learning models from scratch, the patent retrieves and utilizes existing medical images, key findings, and diagnosis names from a database of similar cases. The system copies relevant information from these pre-existing resources to detect lesions in new examination images, avoiding the complexity of training while maintaining reliable detection performance through knowledge transfer from documented cases.
Solution Approach 2:
The patent performs preliminary retrieval and classification of medical images along with their key findings and diagnosis names before actual lesion detection. By pre-organizing the database with similar cases and their associated diagnostic information, the system prepares the necessary knowledge base in advance, eliminating the need for complex real-time machine learning inference and simplifying implementation.
3Adaptability or versatility
If the system retrieves and classifies medical images by similarity, then the ability to detect various lesion types is improved, but the processing time and computational resources increase
Solution Approach 1:
The patent segments the medical image database into distinct groups based on similarity, organ type, and associated key findings. By dividing the large dataset into manageable segments that can be retrieved and classified independently, the system reduces the computational burden of processing all images simultaneously while maintaining comprehensive coverage of various lesion types.
Solution Approach 2:
Instead of processing all medical images in the database, the patent retrieves only the necessary subset of images that are similar to the examination image being analyzed. This partial action approach filters out irrelevant data early in the process, reducing processing time and computational resources while still providing sufficient information for accurate lesion detection across various types.
Data Source
AI summary
A diagnosis support device receives an input retrieval word, retrieves, from a plurality of medical images, each including a lesion region and associated with a key finding that is a key to specify a diagnosis name of the lesion region, a medical image associated with a key finding corresponding to the retrieval word, classifies a retrieved medical image group according to a degree of similarity of images, and detects a lesion candidate region in an examination image on the basis of the degree of similarity between the examination image and a classification result.