Automated Optimal Image Selection in Cardiac MRI Using Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge lies in automating the selection of a single optimal image from a sequence, such as a time series or stack of images, particularly in medical imaging like cardiac MRI, where human expertise is required for selecting inversion times, leading to inconsistencies and increased patient call-backs due to incorrect selections.
Innovation Solution
An ensemble convolutional neural network (CNN) model combines spatial and temporal features using a sliding window approach with a VGG classifier and long short-term memory (LSTM) networks to classify image frames, enabling automated selection of the optimal image without human intervention.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If human expertise is used to select inversion times in cardiac MRI, then diagnostic quality can be maintained, but inconsistencies and increased patient call-backs occur due to subjectivity and variability
Solution Approach 1:
The patent replaces the human mechanical selection process with an automated deep learning system. The neural network model processes image sequences and automatically identifies optimal frames, eliminating human subjectivity and variability in inversion time selection, thereby improving consistency and reducing call-backs
Solution Approach 2:
The system enables self-service automation where the deep learning model independently performs the image selection task without requiring human intervention. The model trains on annotated data and autonomously predicts optimal inversion times, making the process objective and reproducible across different users
2Reliability
If multiple image acquisitions are performed to ensure complete imaging, then imaging completeness is improved, but exam time increases significantly
Solution Approach 1:
The patent applies preliminary action by using deep learning models to predict and identify optimal image frames before final acquisition. The system analyzes image sequences in real-time and pre-determines the best frames for diagnosis, allowing the system to stop early and avoid unnecessary additional acquisitions, thereby reducing exam time while maintaining completeness
Solution Approach 2:
The automated deep learning system replaces the manual, time-consuming process of reviewing multiple acquisitions to determine imaging completeness. The neural network objectively evaluates image quality and diagnostic sufficiency, enabling faster decision-making and reducing the time required to achieve complete imaging
3Manufacturing precision
If specialized technologist expertise is required for cardiac MRI acquisition, then image quality is maintained, but device complexity and cost increase
Solution Approach 1:
The patent replaces the requirement for specialized human expertise with an automated deep learning system. The neural network model, trained on extensive annotated data, performs the complex task of identifying optimal imaging planes and inversion times, eliminating the need for highly specialized technologist training while maintaining image quality standards
Solution Approach 2:
The system creates a digital copy of expert knowledge through the trained neural network model. The model learns from annotated examples and reproduces expert-level image selection capabilities, making the expertise reusable, transferable, and independent of individual human operators, thereby reducing dependency on specialized training
Data Source
AI summary
A method for identification of an optimal image within a sequence of image frames includes inputting the sequence of images into a computer processor configured for executing a plurality of neural networks and applying a sliding window to the image sequence to identify a plurality of image frame windows. The image frame windows are processed using a first neural network trained to classify the image frames according to identified spatial features. The image frame windows are also processed using a second neural network trained to classify the image frames according to identified serial features. The results of each classification are concatenated to separate each of the image frame windows into one of two classes, one class containing the optimal image. An output is generated to display image frame windows classification as including the optimal image.