.png){kind=icon}
.png){kind=icon}
Return
Autoregressive entropy coding with learned contextual prior for lossless medical image compression
.png){kind=icon}
PRE
AI
L
X
R
Y
X
L
DOI：10.1117/1.JEI.35.2.023018.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
To meet diagnostic requirements, medical images such as computed tomography and magnetic resonance imaging typically require lossless compression. However, most existing methods fail to fully exploit prior information in medical images, such as symmetrical structure and contextual dependency. To address this problem, we propose a lossless medical image compression method that combines original image compression with residual image compression. The original image is compressed by the versatile video coding (VVC) test model (VTM) codec, and the reconstructed image and residual image are obtained. For the reconstructed image, we design a convolutional neural network with an attention mechanism, termed CANet, to generate a contextual prior. Within CANet, we introduce a deformable atrous convolution module to extract multiscale semantic features. We also add a structure-aware attention block that integrates a GroupNorm module, coordinate attention, and strip pooling attention to capture context information for adaptive bitrate allocation. For the residual image, a dual checkerboard-decomposition-based autoregressive entropy coding with learned contextual prior is performed to boost compression efficiency. In data training, a weighted cross-entropy loss function is adopted to tackle the unbalanced distribution of residuals. Extensive experiments prove that the proposed method achieves lossless compression at significantly lower bitrates, outperforming recent state-of-the-art baselines.
Keywords:
medical image
lossless compression
entropy coding
contextual prior
checkerboard decomposition
autoregressive model
Journal
J
IF:
1
Papers:
109
Citations:
2.7K
Organization
Cited Papers