Patchwise 3D Bone Reconstruction From Low-Resolution CT
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current osteoporosis diagnosis methods fail to accurately assess bone quality due to limitations in imaging resolution and radiation exposure, leading to misdiagnosis and overdiagnosis, and existing deep learning and topology optimization techniques face challenges in reconstructing high-resolution skeletal images efficiently.
Innovation Solution
A deep learning-based method for patch-wise 3D bone microstructure reconstruction from clinical CT scan data, involving segmentation of low-resolution images into patches, using a pretrained neural network to acquire high-resolution patches, and assembling them with postprocessing to minimize discontinuities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning-based image reconstruction technology is applied to skeletal images, then image resolution can be improved, but training difficulty increases due to complex bone microstructure shapes
Solution Approach 1:
The patent segments the complex bone microstructure into multiple patches, transforming the difficult task of reconstructing entire high-resolution skeletal images into easier sub-tasks of reconstructing individual patches. This segmentation approach reduces training difficulty by presenting simpler, localized structures to the deep learning model while still achieving high overall image resolution.
2Measurement precision
If high-resolution skeletal images are acquired through hardware improvement, then bone microstructure detail is improved, but radiation exposure and imaging time increase
Solution Approach 1:
The patent creates a high-resolution copy of the bone microstructure through deep learning reconstruction from low-resolution input images. Instead of acquiring high-resolution images directly through hardware (which would increase radiation exposure), the system generates a detailed computational copy that preserves bone microstructure information without the harmful side effects of high-dose imaging.
3Measurement precision
If topology optimization-based technology is used for high-resolution reconstruction, then magnification can reach around 10 times, but computational time becomes excessive
Solution Approach 1:
The patent replaces the iterative mechanical optimization process of topology optimization with a deep learning-based computational approach. The trained neural network model directly predicts high-resolution bone microstructure from low-resolution input, eliminating the need for repeated finite element analysis and iterative optimization, thereby dramatically reducing computational time while maintaining high magnification capability.
4Measurement precision
If accurate bone strength evaluation is performed using both BMD measurement and bone microstructure analysis, then diagnostic accuracy is improved, but imaging complexity increases
Solution Approach 1:
The patent creates a universal deep learning model that simultaneously performs multiple functions: it reconstructs high-resolution bone microstructure images and enables bone strength evaluation from a single low-resolution input image. This multi-functional approach achieves accurate diagnostic evaluation (combining BMD and microstructure analysis) without requiring separate imaging procedures, thereby managing imaging complexity through a unified system.
Data Source
AI summary
Provided are a method and device for deep learning-based patch-wise three-dimensional (3D) bone microstructure reconstruction from clinical CT scan data. A computer device may be configured to segment a low-resolution skeletal image into a plurality of low-resolution image patches, to acquire a plurality of high-resolution image patches from the low-resolution image patches, respectively, using a pretrained artificial neural network, and to reconstruct a high-resolution bone microstructure by assembling and postprocessing the high-resolution image patches.