Medical Imaging Apparatus Deep Learning Plane Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current medical imaging technologies face challenges in efficiently extracting a target plane from 2D or 3D image data due to imaging person and target dependency, leading to time-consuming and labor-intensive measurement processes, especially in ultrasonic imaging, where variations in image data affect the accuracy of plane selection.
Innovation Solution
A medical imaging apparatus equipped with a learning model that detects predetermined structures in image data using multi-scale learning data, allowing for step-by-step structure extraction and plane evaluation, thereby reducing computation and processing time while maintaining high accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If template matching is used to determine target planes from 3D data, then plane selection can be automated, but the system cannot cope with imaging person dependency and imaging target dependency, reducing adaptability
Solution Approach 1:
The patent changes the approach from fixed template matching to deep learning-based feature extraction, where the system learns optimal parameters and features from training data. This allows the system to adapt to different imaging conditions, persons, and targets by adjusting its learned parameters rather than relying on predetermined templates.
Solution Approach 2:
The system transitions from static template matching to dynamic deep learning models that can adapt their features and parameters based on input data characteristics. The deep learning model dynamically adjusts to different imaging scenarios, making the plane selection process both automated and adaptable to variations.
2Measurement precision
If deep learning models with high processing capability are used to improve discrimination accuracy, then measurement precision improves, but the hardware requirements and processing time increase
Solution Approach 1:
The patent uses pre-trained deep learning models that have been copied and deployed on medical imaging devices. Instead of requiring complex custom hardware, the system utilizes standardized deep learning frameworks and pre-trained models that can be efficiently executed on existing hardware platforms, reducing device complexity while maintaining high measurement precision.
Solution Approach 2:
The deep learning models are trained in advance on large datasets before deployment. This preliminary training action allows the models to achieve high accuracy without requiring complex hardware during actual operation. The heavy computational work is performed beforehand, and the trained models are then efficiently executed on medical imaging devices.
3Adaptability or versatility
If manual plane selection is performed to account for imaging variations, then adaptability to different conditions is maintained, but the measurement process becomes time-consuming and labor-intensive
Solution Approach 1:
The system implements self-service through automated deep learning-based plane selection that performs the work previously requiring manual inspection. The algorithm independently analyzes imaging data, identifies optimal planes, and makes selections without human intervention, thereby maintaining adaptability while eliminating time loss associated with manual processes.
Solution Approach 2:
The deep learning system incorporates feedback mechanisms where the model continuously learns from training data and improves its plane selection accuracy. The system receives feedback from labeled training examples and adjusts its parameters accordingly, enabling automated adaptation to different imaging conditions without manual intervention during actual measurements.
Data Source
AI summary
Provided is a technology for extracting an image of a target plane from 2D or 3D image data acquired by a medical imaging apparatus with a small amount of computation and at high speed. A plane of a target plane including a predetermined structure is extracted from image data of a subject. A region of the predetermined structure included in the plane is detected by applying a learning model learned using learning data including a target plane for learning including an image of the structure and a region-of-interest plane for learning obtained by cutting out and enlarging a partial region including the structure in the target plane for learning to a plurality of planes obtained from the image data, and the plane of the target plane is extracted based on the detected region of the predetermined structure.