Iterative Sparse Representation for Cross-Modality Medical Image Synthesis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Medical image analysis algorithms face poor performance due to discrepancies between training and testing data distributions, primarily caused by modality heterogeneity and domain variations, such as differences between MRI and CT images, necessitating a mechanism to adapt medical image systems to new modalities without collecting large new data samples.
Innovation Solution
The method employs iterative sparse representation propagation to synthesize target medical images across domains by jointly training source and target dictionaries, enabling cross-modality or cross-domain medical image synthesis, which can be used for training machine learning classifiers and improving anatomical object detection and segmentation tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If medical image analysis algorithms are trained on data from one modality or domain, then they achieve good performance on that specific domain, but they perform poorly when tested on data from different modalities or domains
Solution Approach 1:
The patent synthesizes target domain images by copying and transforming source domain images through iterative sparse representation propagation. The system creates artificial target domain images from source domain images, effectively copying the structural information while transforming it to match the target modality distribution, thereby enabling algorithms to adapt to new domains without retraining
Solution Approach 2:
The patent changes the parameter space by learning separate dictionaries for source and target domains, then using iterative optimization to propagate sparse representations. This parameter transformation allows the same algorithm to handle different modalities by adjusting the dictionary parameters rather than requiring separate trained models for each modality
2Adaptability or versatility
If additional medical images are collected across different modalities to improve training data diversity, then cross-modality performance improves, but the cost and time required for data acquisition increases
Solution Approach 1:
The patent performs preliminary action by pre-training separate source and target dictionaries during an offline phase. These pre-computed dictionaries capture the essential characteristics of each modality and are stored for later use. When new data arrives, the system can quickly synthesize target images using the pre-trained dictionaries without needing to reprocess or retrain on the entire dataset, significantly reducing the time required for adaptation to new modalities
3Adaptability or versatility
If medical images are synthesized across modalities using traditional image processing methods, then some transformation is achieved, but the quality and accuracy of synthesized images is insufficient for training machine learning models
Solution Approach 1:
The patent implements feedback through iterative optimization. In each iteration, the system computes the sparse representation of the source image patch using the source dictionary, then uses this to generate a target image patch. The generated patch is then used to update the sparse representation, and this process repeats for a predetermined number of iterations. This feedback loop continuously refines the synthesized image quality, ensuring that the output images accurately represent the target modality while maintaining the structural information from the source image
Solution Approach 2:
The patent creates composite representations by combining source domain information with target domain characteristics through joint dictionary learning. The source dictionary captures structural information while the target dictionary captures modality-specific characteristics. By combining these two dictionaries in the sparse representation formulation, the system generates synthesized images that possess both the structural accuracy of the source image and the modality characteristics of the target image, achieving high synthesis quality
Data Source
AI summary
A method and apparatus for medical image synthesis across image modalities or domains is disclosed, which synthesizes a target medical image based on a source medical image. A plurality of image patches are cropped from the source medical image. A synthesized target medical image is then generated from the source medical image by jointly performing sparse coding between each image patch of the source medical image and a corresponding image patch of the synthesized target image based on jointly trained source and target dictionaries.