3D Medical Image Compression With Residual Slice Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing 3D medical image compression methods fail to effectively exploit bilateral correlations within 3D medical images, limiting their adaptability and efficiency.
Innovation Solution
A method that partitions 3D medical images into slices, applies lossy encoding to slices, computes residues, and uses lossless encoding for residues, leveraging intra-slice and inter-slice contexts through bi-directional cross-attention and symmetry-based context extraction to reduce redundancy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional 3D transform-based compression methods are used, then compression is achieved, but adaptability to specific characteristics of 3D medical images is limited
Solution Approach 1:
The patent changes the fundamental parameter of the compression approach from traditional transform-based methods to deep learning-based methods. Specifically, it employs 3D convolutional neural networks and entropy coding models that adapt to the statistical characteristics of medical images, achieving better compression ratios while maintaining diagnostic quality
Solution Approach 2:
The patent replaces traditional mechanical/mathematical transform operations with neural network-based operations. The deep learning model learns optimal transformation and compression strategies from training data, substituting hand-crafted transforms with adaptive, data-driven mechanisms that better capture medical image characteristics
2Quantity of substance
If sequence-based compression methods are used, then some redundancy reduction is achieved, but inter-slice redundancy in 3D medical images is not effectively reduced
Solution Approach 1:
The patent transitions from 2D slice-based compression to 3D volumetric compression by processing multiple slices together through 3D convolutional operations. This dimensional extension enables the model to capture correlations across adjacent slices, effectively reducing inter-slice redundancy while improving compression efficiency
Solution Approach 2:
The patent segments the 3D medical image into multiple slices for parallel processing while maintaining inter-slice relationships through 3D contextual modeling. Each slice is processed individually by the neural network but with access to contextual information from neighboring slices, enabling both efficient computation and effective redundancy reduction
3Productivity
If lossy compression is applied to 3D medical images, then compression efficiency is improved, but image quality and diagnostic accuracy may deteriorate
Solution Approach 1:
The patent changes the compression parameter from lossy to lossless by using entropy coding models that preserve all original image information. The neural network learns optimal probability distributions for entropy coding without discarding any diagnostic information, achieving high compression efficiency while maintaining full diagnostic fidelity
Solution Approach 2:
The patent incorporates feedback mechanisms where the decoded image is compared with the original to compute residuals, which are then encoded. This feedback loop ensures that all lost information from the initial compression is recovered, achieving lossless compression while maintaining high compression ratios through intelligent residual coding
Data Source
AI summary
A method for compressing a 3D medical image includes the steps of receiving a 3D medical image, partitioning the 3D medical image into a plurality of first slices, encoding the plurality of the first slices by a lossy codec into first bitstreams, decoding the first bitstreams by the lossy codec to obtain a plurality of second slices, computing a plurality of residues by comparing the plurality of the first slices and the plurality of the second slices, encoding the plurality of the residues by a lossless codec to obtain a plurality of encoded residues, and outputting the first bitstreams and the plurality of the encoded residues as compressed image data. Each residue corresponds to one of the first slices and its corresponding second slice. Experimental results on prevailing 3D medical image datasets demonstrate that the proposed method achieves promising compression performance and outperforms state-of-the-art methods.