Recent achievements of sequence prediction models in numerous domains, including compression, provide great potential for novel learning-based codecs. In such models, the input sequence’s shape and size play a crucial role in learning the mapping function of the data distribution to the target output. This work examines numerous input configurations and sampling schemes for a many-to-one sequence prediction model, specifically for compressing 3D medical images (16-bit depth) losslessly. The main objective is to determine the optimal practice for enabling the proposed Long Short-Term Memory (LSTM) model to achieve high compression ratio and fast encoding–decoding performance.Our LSTM models are trained with 4-fold cross-validation on 12 high-resolution CT dataset while measuring model’s compression ratios and execution time. Several configurations of sequences have been evaluated, and our results demonstrate that pyramid-shaped sampling represents the best trade-off between performance and compression ratio (up to 3×). We solve a problem of non-deterministic environments that allow our models to run in parallel without much compression performance drop.Experimental evaluation was carried out on datasets acquired by different hospitals, representing different body segments, and distinct scanning modalities (CT and MRI). Our new methodology allows straightforward parallelisation that speeds-up the decoder by up to 37× compared to previous methods. Overall, the trained models demonstrate efficiency and generalisability for compressing 3D medical images losslessly while still outperforming well-known lossless methods by approximately 17% and 12%. To the best of our knowledge, this is the first study that focuses on voxel-wise predictions of volumetric medical imaging for lossless compression.