Deep Learning-based Volumetric Compression For High Resolution Brain Image

With the rapid development of biomicroscopic imaging technologies,volumetric whole brain imaging techniques at single-neural resolution make great progress.They produce high-resolution images and the images show soma,axon and synapse clearly,which is significant for the study on neuron morphology and structure and also facili-cates the development of artificial intelligence.However,such high resolution volumet-ric images bring great challenge for storage and transmission,so looking for efficient compression methods for volumetric brain image has become an urgent task.Nowadays many medical image compression methods adopt wavelet-transform-based techniques,where JP3D method is volumetric extension of JPEG2000 standard and it can compress volumetric high-bit-depth image efficiently.However,this method has two drawbacks.Firstly,the decoded images of this method have distortion effects,and the higher compression rates will cause more obvious distortion effects.These dis-tortion effects will affect the study on brain images.Secondly,the EBCOT entropy coding algorithm of JP3D doesn't exploit the correlations between subbands,which causes low compression efficiency.Recently,the 2D natural image compression meth-ods based on deep learning achieves outstanding compression performance,and its com-pression efficiency outperforms conventional image compression methods,e.g.BPG,JPEG2000,JPEG.Although deep learning-based 2D image compression achieves great progress,there is rare study on 3D deep learning-based image compression.This paper aims to improve compression efficiency of high resolution volumetric brain images,and to obtain higher quality decoded images at the same bit rates.In-spired by 2D deep learning-based image compression,the paper focuses on two aspects research:1.We study the improvement methods of JP3D based on deep learning.On the one hand,we propose a volumetric post-processing neural network for 3D brain image compression to reduce compression distortion effects.We adopt convolutional neural network to capture features on each dimension of volumetric image,and meanwhile we introduce 3D residual block to accelerate network convergence.The experiment results demonstrate that this method reduces the compression effects of compressed brain images and it also performs well on volumetric high-dynamic-range images.On the other hand,we combine entropy model based on 3D pixelCNN and RNN with JP3D method,and exploit correlations between 3D coefficient subbands of wavelet transform to improve entropy coding efficiency.2.We study volumetric image compression method based on deep learning.This paper proposes a volumetric end-to-end image compression method based on deep learning.The method transforms the volumetric input image into lower-dimensional latent space by auto-encoder,and then uses hyperprior model and conditional context model as entropy model to jointly estimate the probability distribution of latent space.Meanwhile,we introduce 3D non-local attention models to exploit global correlations of input image,and we also adopt 3D ConvLSTM based context model to reduce model complexity,which encodes the latent features sequentially along channel dimension to exploit causal correlations between channels.Each model in the compression pipeline is optimized through a rate-distortion function.Experiment results demonstrate that our proposed volumetric image compression method has a significant advantage over state-of-the-art 2D image compression method based on deep learning,and our method outperforms JP3D and HEVC method,especially at lower bit rates.