Image Registration via DCT Residual Maps for Digital Pathology
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image registration methods for digital pathology face challenges in efficiently aligning large, high-resolution digital slides with non-linear deformations, leading to slow processing times and reduced accuracy, especially in virtual microscopy applications.
Innovation Solution
A method is introduced that determines an image distortion map by selecting low-resolution representations of images, calculating current and residual distortion maps using discrete cosine transforms, and iteratively refining the alignment process to improve registration accuracy and speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional image registration methods are used to align digital slides with non-linear deformations, then registration accuracy is maintained, but processing time increases significantly
Solution Approach 1:
The patent divides the large digital slide into multiple smaller image tiles or patches. Each tile is registered independently using traditional accurate methods, then the results are combined. This segmentation allows accurate registration to be applied to manageable portions rather than the entire large image, reducing overall processing time while maintaining accuracy through local detailed analysis.
Solution Approach 2:
The patent implements a multi-scale approach where a coarse alignment is first established using down-sampled or low-resolution versions of the images. This preliminary rough registration provides an initial transformation that brings the images into approximate alignment, after which finer detailed registration is performed on the pre-aligned images. This preliminary action reduces the search space for subsequent accurate registration, significantly speeding up the process.
2Measurement precision
If high-resolution images are processed directly for registration, then registration accuracy is improved, but computational complexity and processing time increase
Solution Approach 1:
The patent segments the high-resolution images into smaller tiles that can be processed independently with reduced computational requirements. Each tile requires less memory and computational power to register accurately, while the overall registration accuracy is maintained through proper tiling and boundary handling strategies.
Solution Approach 2:
The patent performs preliminary down-sampling to create lower-resolution versions of the high-resolution images. These down-sampled images are used to establish initial alignment parameters and guide the subsequent high-resolution registration process. This preliminary action at reduced resolution significantly lowers computational complexity before the final accurate registration is performed.
Solution Approach 3:
The patent introduces a resolution dimension by working across multiple scales or resolutions. Instead of directly registering high-resolution images, the method operates across a hierarchy of resolutions, using lower resolutions for coarse alignment and progressively refining to higher resolutions. This dimensional approach to resolution management reduces computational complexity while preserving registration accuracy.
Data Source
AI summary
A method for registering images aligns a fixed image with a corresponding moving image. A target image and a reference image are selected. The reference and target images are relatively low-res representations of the fixed and moving images respectively. A current image distortion map is determined that defines distortions between the target and reference images. A residual image is determined by comparing the reference image and a warped target image derived from the target image based on the current image distortion map. A residual image distortion map is determined based on transform coefficients of cosine functions fitted to the residual image. The coefficients are determined by applying a DCT to a signal formed by the residual image and image gradients of the warped target or reference image. The current image distortion map is combined with the residual image distortion map to align the fixed image and the moving image.