ML Medical Image Reconstruction for Full-FOV Attenuation Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Nuclear imaging systems face challenges in accurately correcting non-attenuated corrected (NAC) images due to limited field-of-view (FOV) attenuation maps generated from CT images, leading to inaccuracies in medical metrics like organ volume computation and comparison across scans.
Innovation Solution
Employing machine learning processes to extend partial CT images to full FOV and generate corresponding attenuation maps, using additional images like x-ray and optical images to correct NAC images, and create segmentation masks for accurate organ identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a CT image is scanned with a limited field-of-view to reduce scan time and radiation dose, then scan efficiency is improved, but the attenuation map coverage is insufficient leading to inaccurate attenuation correction
Solution Approach 1:
The patent creates a synthetic full-FOV CT image by copying and extending information from the partial-FOV CT image using deep learning. The neural network synthesizes the missing anatomical regions by learning from the available partial image and training data, generating a complete attenuation map without requiring additional scanning.
Solution Approach 2:
The patent transforms the field-of-view parameter from limited to full coverage through computational methods. By changing the spatial coverage parameter using deep learning synthesis, the system achieves complete attenuation map coverage while maintaining the original limited scan parameters for efficiency.
2Measurement precision
If a CT image covers the entire patient to ensure complete attenuation correction, then attenuation correction accuracy is improved, but scan time and radiation dose increase
Solution Approach 1:
Instead of physically scanning the entire patient, the system copies and synthesizes the missing regions computationally. The deep learning model generates synthetic image data for unscanned regions, creating a complete attenuation map from a partial physical scan.
Solution Approach 2:
The system performs preliminary synthesis of the complete attenuation map before attenuation correction is needed. By pre-computing the full FOV image from the partial scan using deep learning, the system prepares complete correction data in advance without extending the actual scan time.
3Measurement precision
If a CT image covers the entire patient to ensure complete organ coverage, then medical metrics accuracy is improved, but scan time and radiation dose increase
Solution Approach 1:
The system synthesizes complete organ coverage by copying and extending anatomical information from the partial scan. The deep learning model generates synthetic representations of organs in unscanned regions, enabling complete organ volume and contribution percentage calculations.
4Measurement precision
If multiple machine learning processes are applied to extend FOV and generate attenuation maps, then image reconstruction accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent combines multiple machine learning functions into an integrated deep learning framework. The system merges FOV extension, segmentation, and attenuation map generation into a unified neural network architecture that processes all tasks simultaneously, reducing overall computational complexity compared to sequential processing.
Data Source
AI summary
Machine learning based systems and methods for improving attenuation and scatter correction and for estimating organ volume in medical images are disclosed. In some embodiments, a co-modality image, a non-attenuated corrected nuclear image, and an x-ray image are received. A first machine learning process is applied to the co-modality image to generate location data identifying feature locations. Further, a second machine learning process is applied to the non-attenuated corrected nuclear image, the x-ray image, and the location data to generate a segmentation mask. In addition, at least a third machine learning process is applied to the segmentation mask, the co-modality image, the non-attenuated corrected nuclear image, and the x-ray image to generate a plurality of synthetic images. Moreover, at least a fourth machine learning process is applied to the plurality of synthetic images to generate a final synthetic image.