3D Medical Image Lossless Compressor Using Deep Learning Approaches

Abstract

The ever-increasing importance of accelerated information processing, communica-tion, and storing are major requirements within the big-data era revolution. With the extensive rise in data availability, handy information acquisition, and growing data rate, a critical challenge emerges in efficient handling. Even with advanced technical hardware developments and multiple Graphics Processing Units (GPUs) availability, this demand is still highly promoted to utilise these technologies effectively. Health-care systems are one of the domains yielding explosive data growth. Especially when considering their modern scanners abilities, which annually produce higher-resolution and more densely sampled medical images, with increasing requirements for massive storage capacity. The bottleneck in data transmission and storage would essentially be handled with an effective compression method. Since medical information is critical and imposes an influential role in diagnosis accuracy, it is strongly encouraged to guarantee exact reconstruction with no loss in quality, which is the main objective of any lossless compression algorithm. Given the revolutionary impact of Deep Learning (DL) methods in solving many tasks while achieving the state of the art results, includ-ing data compression, this opens tremendous opportunities for contributions. While considerable efforts have been made to address lossy performance using learning-based approaches, less attention was paid to address lossless compression. This PhD thesis investigates and proposes novel learning-based approaches for compressing 3D medical images losslessly.Firstly, we formulate the lossless compression task as a supervised sequential prediction problem, whereby a model learns a projection function to predict a target voxel given sequence of samples from its spatially surrounding voxels. Using such 3D local sampling information efficiently exploits spatial similarities and redundancies in a volumetric medical context by utilising such a prediction paradigm. The proposed NN-based data predictor is trained to minimise the differences with the original data values while the residual errors are encoded using arithmetic coding to allow lossless reconstruction.Following this, we explore the effectiveness of Recurrent Neural Networks (RNNs) as a 3D predictor for learning the mapping function from the spatial medical domain (16 bit-depths). We analyse Long Short-Term Memory (LSTM) models’ generalisabil-ity and robustness in capturing the 3D spatial dependencies of a voxel’s neighbourhood while utilising samples taken from various scanning settings. We evaluate our proposed MedZip models in compressing unseen Computerized Tomography (CT) and Magnetic Resonance Imaging (MRI) modalities losslessly, compared to other state-of-the-art lossless compression standards.This work investigates input configurations and sampling schemes for a many-to-one sequence prediction model, specifically for compressing 3D medical images (16 bit-depths) losslessly. The main objective is to determine the optimal practice for enabling the proposed LSTM model to achieve a high compression ratio and fast encoding-decoding performance. A solution for a non-deterministic environments problem was also proposed, allowing models to run in parallel form without much compression performance drop. Compared to well-known lossless codecs, experimental evaluations were carried out on datasets acquired by different hospitals, representing different body segments, and have distinct scanning modalities (i.e. CT and MRI).To conclude, we present a novel data-driven sampling scheme utilising weighted gradient scores for training LSTM prediction-based models. The objective is to determine whether some training samples are significantly more informative than others, specifically in medical domains where samples are available on a scale of billions. The effectiveness of models trained on the presented importance sampling scheme was evaluated compared to alternative strategies such as uniform, Gaussian, and sliced-based sampling.