Deep Learning Image Reconstruction via Forward Projection Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current nuclear imaging systems using deep learning for image reconstruction often suffer from algorithmic biases leading to errors such as hallucinations, which can result in false confidence and subpar or misdiagnosis.
Innovation Solution
The system employs a computer-implemented method that includes receiving input projection data, applying a machine learning process to generate output image data, applying a forward projection process to determine a loss value, and training the machine learning process based on this loss value. The trained model is then used to reconstruct medical images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If deep learning algorithms are used for image reconstruction, then image reconstruction speed and quality are improved, but algorithmic biases lead to hallucinations and diagnostic errors
Solution Approach 1:
The patent implements a feedback mechanism where the reconstructed image is forward-projected back to projection space and compared with the original measured projection data. This feedback loop calculates a loss value that quantifies the discrepancy, which is then used to iteratively adjust the deep learning model's parameters, ensuring the reconstructed image remains faithful to the actual measured data and prevents hallucinations
Solution Approach 2:
The patent performs preliminary forward projection of the reconstructed image before final diagnostic use. By forward-projecting the reconstructed image back to projection space and comparing it with the original measured data, the system proactively identifies and corrects potential hallucinations before they reach the diagnostic stage
2Reliability
If traditional mathematical algorithms are used for reconstruction, then algorithmic reliability is maintained, but reconstruction quality and speed are limited
Solution Approach 1:
The patent merges traditional mathematical reconstruction algorithms with deep learning-based iterative reconstruction. The system combines the reliability of conventional algorithms with the quality enhancement of deep learning by using the deep learning model to generate initial reconstructions while incorporating traditional forward projection and loss calculation to ensure diagnostic accuracy
3Productivity
If deep learning models are trained without forward projection validation, then training speed is improved, but model accuracy and trustworthiness deteriorate
Solution Approach 1:
The patent incorporates forward projection validation as a feedback mechanism during the training process. The loss function compares forward-projected reconstructed images with original measured projection data, providing continuous feedback that guides the model training toward greater accuracy and trustworthiness while maintaining reasonable training efficiency
Data Source
AI summary
Systems and methods for training end-to-end deep learning reconstruction processes, and for reconstructing medical images based on the trained deep learning processes, are disclosed. In some examples, input projection data is received. An untrained machine learning process is applied to the input projection data and, based on the application of the machine learning process to the projection data, an output image is generated. Further, a forward projection process is applied to the output image and, based on the application of the forward projection process to the output image, forward projected image data is generated. A loss value is then determined based on the forward projected image data and the input projection data. The loss value is then compared to a threshold value to determine whether the machine learning process is trained. The trained machine learning process may be employed to reconstruct images, such as positron emission tomography (PET) images.