Synthetic Radiologic Image Generation With Slice-Based Diffusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing 2D and 2.5D diffusion models in medical imaging suffer from stripe and streaking artifacts in sagittal and coronal views, and 3D denoising diffusion models face high computational resource requirements and overfitting risks due to limited 3D volume availability.
Innovation Solution
A trained conditional generative model generates synthetic radiologic images using a series of seeds, combining them to reduce artifacts and optimize computational efficiency, employing diffusion models like Latent Diffusion Models (LDMs) and Denoising Diffusion Implicit Models (DDIMs) with conditional inputs and image embeddings.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If 3D denoising diffusion probability models or 3D latent diffusion models are used directly, then stripe and streaking artifacts are reduced, but computational resource requirements increase significantly
Solution Approach 1:
The patent segments the 3D volume into multiple 2D axial slices for independent processing by the diffusion model. Instead of processing the entire 3D volume at once, each slice is generated separately and then assembled into the complete 3D reconstruction, significantly reducing computational memory requirements while maintaining diagnostic quality
Solution Approach 2:
The patent transforms the 3D reconstruction problem into a series of 2D generation problems. By processing axial slices in 2D space and then reconstructing the 3D volume through assembly, the method achieves efficient computation while preserving three-dimensional anatomical information
2Loss of information
If 3D volumes are used for training autoencoders and diffusion models, then comprehensive anatomical information is captured, but the scarcity of available 3D volumes poses the risk of overfitting
Solution Approach 1:
The patent enables training on 2D axial slices instead of requiring scarce 3D volumes. This dimensionality transformation expands the available training data pool significantly, allowing models to learn from numerous 2D slices while maintaining the ability to reconstruct 3D anatomy, thereby preventing overfitting
3Productivity
If 2D or 2.5D diffusion models are used, then computational efficiency is improved, but stripe or streaking artifacts occur in sagittal and coronal views
Solution Approach 1:
The patent processes the 3D volume by segmenting it into multiple axial slices that are independently generated by the diffusion model. This segmentation approach maintains computational efficiency while the subsequent assembly process ensures complete anatomical coverage without artifacts
Solution Approach 2:
The patent merges multiple independently generated axial slices into a complete 3D volume reconstruction. This combining process restores the full anatomical information and eliminates artifacts by integrating information from all slices coherently
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
Systems, methods, and computer programs disclosed herein relate to generating synthetic radiologic images.