Lesion Detection in Medical Imaging Using Dual Neural Network Identifiers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current medical image processing technologies face challenges in accurately distinguishing between diffusive lung disease lesions and blood vessels, particularly in three-dimensional images with thick slice thicknesses, where collecting correct data for machine learning is difficult.
Innovation Solution
A medical image processing apparatus and method using two identifiers, one for two-dimensional and one for three-dimensional image analysis, both employing convolutional neural networks to identify lesion regions and differentiate between blood vessels and bronchial regions, with the first identifier handling thick slice thicknesses and the second identifier utilizing thinner slices for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single three-dimensional identifier is used to detect lesion regions, then the detection can be performed in 3D space, but the accuracy is insufficient to distinguish between blood vessels and bronchial regions
Solution Approach 1:
The patent divides the identification task into two separate identifiers: a first identifier for detecting lesion region candidates and a second identifier for classifying them as blood vessel or bronchial region. This segmentation allows each identifier to specialize in its specific function, improving overall detection accuracy and reliability.
Solution Approach 2:
The first identifier acts as an intermediary that generates lesion region candidates from the medical image, which are then passed to the second identifier for classification. This intermediary step enables the system to handle the complex distinction between blood vessels and bronchial regions more effectively.
2Ease of manufacture
If machine learning is performed using thick slice thickness images, then practical clinical data can be collected, but the ability to distinguish fine structures like blood vessels deteriorates
Solution Approach 1:
The patent changes the parameter of slice thickness used for training different identifiers. The first identifier is trained with thick slice images (5mm or more) that are readily available in clinical practice, while the second identifier is trained with thin slice images (1mm or less) that provide fine structural detail. This parameter change allows each identifier to optimize for its specific function.
3Productivity
If only two-dimensional image analysis is used, then processing is simple and fast, but the ability to detect three-dimensional lesion structures deteriorates
Solution Approach 1:
The patent transitions from two-dimensional image analysis to three-dimensional analysis by introducing a second identifier that performs classification in the depth dimension. The first identifier processes 2D images efficiently, while the second identifier adds 3D contextual understanding to distinguish blood vessels from bronchial regions, achieving both speed and accuracy.
Data Source
AI summary
The medical image processing apparatus includes a medical image acquisition unit that acquires a medical image; and a lesion detection unit that detects a lesion region in the medical image. The lesion detection unit includes a first identifier that identifies a lesion region candidate in the medical image and a second identifier that identifies whether the lesion region candidate identified by the first identifier is a blood vessel region, and detects the lesion region candidate that is not identified as the blood vessel region by the second identifier as the lesion region.