Region-Based Retinal Image Compression Using Convolutional Auto-Encoder And Arithmetic Encoding

In medical imaging, image compression is used to ease bandwidth and storage requirements. In this paper, we report a Roi based image compression approach for retinal images based on convolutional auto-encoder (CAE) and arithmetic coding, to achieve a better compression efficiency with high quality reconstruction. There are three main sections in our compression scheme: preprocessing of data, CAE-based lossy compression, and arithmetic coding for lossless compression. For data preprocessing, an automatic segmentation is used to extract the ROI and non-ROI parts. The CAE is used to achieve a high compression on the non-ROI, the encoding network learns to produce an optimal compacted representation that maintains the structural information. Using the compact representation, the decoding network reconstructs the decoded image. The ROI portion of the image is losslessly compressed using arithmetic coding to preserve diagnostically important information. In order to validate the performance of our proposed approach, we compared our results with state-of-the-art compression techniques and recent methods. At high compression ratio as 40, the proposed approach achieved a mean of 39.5 dB at signal-to-noise ratio (PSNR) metric and $\mathbf{9 7. 2 8 \%}$ similarity in multi-scale structural similarity (MS-SSIM), proving its efficacy.