Practical Lossless Volumetric Medical Image Compression via Tri-plane Context Tree Learning

Yuanchao Bai; Yifan Zhao; Kai Wang; Yuanbo Du; Jie Cheng; Teng Fang; Xianming Liu; Wen Gao

eess.IVAug 14, 2026

Practical Lossless Volumetric Medical Image Compression via Tri-plane Context Tree Learning

Authors: Yuanchao Bai, Yifan Zhao, Kai Wang, Yuanbo Du, Jie Cheng, Teng Fang, Xianming Liu, Wen Gao

Organizations: Faculty of Computing, Harbin Institute of Technology, Harbin, 150001, China · Huawei Tech. Company, Ltd. · School of Electronics Engineering and Computer Science, Peking University, Beijing, 100871, China, and Pengcheng Laboratory, Shenzhen, 518055, China

Abstract

Lossless compression of volumetric medical images is of paramount importance for clinical and research applications where data fidelity is essential. Traditional compression methods are often limited in efficiency due to rigid, handcrafted models. Conversely, deep neural network (DNN)-based compression methods, while effective, demand substantial computational resources, hindering deployment in resource-constrained settings. To address these challenges, we propose a novel tri-plane context tree (TCT)-based method for lossless volumetric medical image compression that delivers high performance without relying on DNNs or external training data. To exploit intra-slice and inter-slice redundancies, we introduce a compact tri-plane context representation that decomposes complex 3D context modeling into efficient 2D modeling on three orthogonal planes. By integrating this representation with a context tree framework, we develop an input-specific TCT model employing an adaptive binary tree structure. At each tree node, the model dynamically selects from a suite of tri-plane based predictors and contextual feature extractors, enabling data-adaptive context modeling tailored to local structural characteristics. Instead of offline training, we sample a subset of the input volume to learn the TCT model by optimizing the minimum description length (MDL) through iterative construction and pruning. With the learned TCT model, each pixel retrieves its corresponding context, computes the prediction residual using the predictor dictated by the context, and performs entropy encoding based on the associated histograms. Experimental results demonstrate that the proposed method achieves compression performance on par with recent DNN-based methods on multiple datasets, while maintaining low computational cost and fast coding speeds, making it highly applicable in practice.

Figures & tables

Notation	Description
${\mathbf{I}}$	volumetric medical image
${\mathbf{I}}_{idx}$	pixel set at index $idx\in\{<i,{\mathcal{S}},{\mathcal{P}},\ldots\}$
$I_{idx}$	pixel at index $idx\in\{i,L[0],L[1],\ldots\}$
${\mathbf{C}}$ , $C_{j}$	context set ${\mathbf{C}}=\{C_{1},\ldots,C_{M}\}$ , context $C_{j}\in{\mathbf{C}}$
${\mathrm{pred}}$ , ${\mathrm{pred}}_{j}$	predictor and predictor assigned to $C_{j}$
${\mathbf{r}}_{idx}$	residual set at index $idx$

TABLE I: Notations for TCT model and TCT-based compression method.

Fig. 1: Example of a volumetric medical image and its tri-plane representation, including x-y, y-z, x-z planes.

Panel of Figure 3 — Fig. 2: Illustration of tri-plane context representation for encoding pixel $I_{i}$ . (a) all previously encoded pixels ${\mathbf{I}}_{<i}$ . (b) x-y context plane. (c) y-z context plane. (c) x-z context plane. Red: the pixel $I_{i}$ . Blue: the previously encoded pixels. Orange: the pixels in each context plane.

Fig. 3: Diagram of the tri-plane context tree (TCT) model. Each internal node is assigned a feature index $k$ corresponding to the $k$ -th feature $f^{k}_{i}$ ’s, which are derived from a combination of pixels (marked with orange) residing on the three context planes. The internal node is also associated with a threshold $t$ for binary partitioning. Each leaf node represents a distinct context. For each pixel, the TCT model is traversed from the root to the leaf based on the pixel’s features, retrieving the corresponding context.

Fig. 4: Block diagram of the TCT-based volumetric medical image compression. The TCT-based compression method requires no offline training. Instead, it learns a customized, input-specific context model ${\mathcal{T}}$ by performing TCT learning on a sample set ${\mathcal{S}}$ of the input 3D volume. During compression, each pixel retrieves its corresponding context from the learned model ${\mathcal{T}}$ to enable adaptive prediction and entropy coding. Finally, both the learned TCT model and the compressed image data are stored into the bitstream to ensure decodablity.

Fig. 5: Labeling of neighboring pixels used in prediction and contextual feature extraction on three context planes. The x-y, y-z and x-z context planes enjoy the same context pattern. The $\bigtriangleup$ denotes the current pixel. The $L$ denotes the left pixel of $\bigtriangleup$ . The $LL$ denotes the left-left pixel of $\bigtriangleup$ . The $T$ denotes the top pixel of $\bigtriangleup$ . The $TT$ denotes the top-top pixel of $\bigtriangleup$ . The $LT$ denotes the left-top pixel of $\bigtriangleup$ . The $LLT$ denotes the left-left-top pixel of $\bigtriangleup$ . The $RT$ denotes the right-top pixel of $\bigtriangleup$ .

Predictor	Sub-predictor
Median-2D	${\mathrm{median}}(I_{L[0]},I_{L[0]}+I_{T[0]}-I_{LT[0]},I_{T[0]})$
Median-3D	${\mathrm{median}}(I_{L[0]},I_{L[0]}+I_{T[0]}-I_{LT[0]},I_{T[0]})$
	${\mathrm{median}}(I_{L[1]},I_{L[1]}+I_{T[1]}-I_{LT[1]},I_{T[1]})$
	${\mathrm{median}}(I_{L[2]},I_{L[2]}+I_{T[2]}-I_{LT[2]},I_{T[2]})$
Blending-3D	$I_{L[0]}-I_{T[0]}+I_{RT[0]}$
	$I_{L[0]}+I_{T[0]}-I_{LT[0]}$

TABLE II: Tri-plane based predictors and the corresponding sub-predictors

Fig. 6: Quadruple representation $(zbit,sbit,msb,lsb)$ of prediction residual.

Type	Feature Value
Gradient-2D	$I_{L[0]}-I_{LT[0]}$
	$I_{LT[0]}-I_{T[0]}$
	$I_{T[0]}-I_{RT[0]}$
	$I_{T[0]}-I_{TT[0]}$
	$I_{L[0]}-I_{LL[0]}$
	$I_{LL[0]}+I_{LT[0]}-I_{LLT[0]}-I_{L[0]}$

TABLE III: Tri-plane based features and the corresponding values

Fig. 7: Illustration of the proposed two-step TCT learning algorithm.

	Chaos	MosMedData	Trabit	MRNet
Type	Abdomen CT	Lung CT	Brain MRI	Knee MRI
Training Set †	20	125	70	1130
Test Set	20	45	30	120
Bit Depth	16	16	16	8
Slice Count	78 $\sim$ 294	32 $\sim$ 72	176	17 $\sim$ 61
Resolution	512 $\times$ 512	512 $\times$ 512	176 $\times$ 208	256 $\times$ 256

TABLE IV: Overview of volumetric medical image datasets in experiments. † Training sets are for DNN-based methods.

Codec	Chaos		MosMedData		Trabit		MRNet
Codec	BPP $\downarrow$	Compression Ratio $\uparrow$	BPP $\downarrow$	Compression Ratio $\uparrow$	BPP $\downarrow$	Compression Ratio $\uparrow$	BPP $\downarrow$	Compression Ratio $\uparrow$
Traditional
PNG	7.09	2.256	7.09	2.257	3.08	5.190	4.56	1.755
JPEG-LS [ 51 ]	5.08	3.148	5.02	3.188	2.22	7.194	4.06	1.971
CALIC [ 54 ]	-	-	-	-	-	-	3.93	2.036
JPEG2000 [ 42 ]	5.05	3.169	5.32	3.009	2.58	6.204	4.17	1.920

TABLE V: Lossless image compression performance of our TCT-based volumetric medical image compression method, compared with other lossless compression methods on Chaos, MosMedData, Trabit and MRNet datasets. The best performance is marked in bold , the second-best is marked with an underline , and the third-best is marked with a double-underline . † Using the released CT compression model trained on MosMedData dataset.

Codec	GPU Max Usage	Chaos		MosMedData
Codec		Enc.T.	Dec.T.	Enc.T.	Dec.T.
Traditional (CPU)
JPEG-LS	-	0.02s	0.02s	0.03s	0.03s
JPEG-XL	-	2.18s	0.05s	4.08s	0.07s
JP3D	-	0.04s	0.04s	0.07s	0.07s
HEVC-RExt	-	4.01s	0.02s	5.61s	0.03s

TABLE VI: Average coding time per slice of volumetric medical image compression methods. The DNN-based methods run on GPU with maximum GPU usage reported.

M-2D	M-3D	B-3D	Chaos	MosMedData	Trabit	MRNet
$\checkmark$	$\times$	$\times$	4.64	4.72	2.00	3.83
$\checkmark$	$\checkmark$	$\times$	4.44	4.71	1.97	3.83
$\checkmark$	$\checkmark$	$\checkmark$	4.27	4.64	1.91	3.79

TABLE VII: Effects of Predictors on the compression performance. M-2D, M-3D, B-3D denote Median-2D, Median-3D and Blending-3D.

Panel of Figure 15 — Fig. 8: The percentage of predictor and feature selections in our TCT-based compression method on the Chaos dataset. (a) predictor selection. (b) feature selection. The M-2D, M-3D and B-3D denote the Median-2D, Median-3D and Blending-3D respectively.

Panel of Figure 16 — Fig. 9: Visualization of the predictor selection map generated by our TCT-based compression method for a 2D slice of a 3D CT scan. (a) 16-bit CT slice mapped to 8-bit for visualization. (b) predictor selection map. Red: Median-2D. Yellow: Median-3D. Blue: Blending-3D.

G-2D	G-3D	Pred	Chaos	MosMedData	Trabit	MRNet
$\checkmark$	$\times$	$\times$	4.36	4.67	1.93	3.86
$\checkmark$	$\checkmark$	$\times$	4.30	4.66	1.93	3.85
$\checkmark$	$\checkmark$	$\checkmark$	4.27	4.64	1.91	3.79

TABLE VIII: Effects of features on compression performance. G-2D, G-3D, Pred denote Gradient-2D, Gradient-3D and Prediction.

Percentage	Chaos	MosMedData	Trabit	MRNet
Model	0.04%	0.08%	0.11%	0.14%
Image Data	99.96%	99.92%	99.89%	99.86%

TABLE IX: Average proportions of TCT model and compressed image data.

TCT Pruning	Histogram Num.		Comp. Performance
TCT Pruning	$\times$	$\checkmark$	$\times$	$\checkmark$
Chaos	4048	224	4.28	4.27
MosMedData	2552	198	4.67	4.64
Trabit	691	106	1.93	1.91
MRNet	587	111	3.82	3.79

TABLE X: Effects of TCT pruning on average number of histograms and compression performance.

Fig. 10: Effects of $(\lambda,u)$ on average lossless compression performance (bpp) on random samples from Chaos, MosMedData, Trabit and MRNet datasets. The (50, 224) is selected for all four datasets.

Explore similar work

Mar 12, 2026

MedPruner: Training-Free Hierarchical Token Pruning for Efficient 3D Medical Image Understanding in Vision-Language Models

While specialized Medical Vision-Language Models (VLMs) have achieved remarkable success in interpreting 2D and 3D medical modalities, their deployment for 3D volumetric data remains constrained by significant computational inefficiencies. Current architectures typically suffer from massive anatomical redundancy due to the direct concatenation of consecutive 2D slices and lack the flexibility to handle heterogeneous information densities across different slices using fixed pruning ratios. To address these challenges, we propose MedPruner, a training-free and model-agnostic hierarchical token pruning framework specifically designed for efficient 3D medical image understanding. MedPruner introduces a two-stage mechanism: an Inter-slice Anchor-based Filtering module to eliminate slice-level temporal redundancy, followed by a Dynamic Information Nucleus Selection strategy that achieves adaptive token-level compression by quantifying cumulative attention weights. Extensive experiments on three 3D medical benchmarks and across three diverse medical VLMs reveal massive token redundancy in existing architectures. Notably, MedPruner enables models such as MedGemma-1.5 to maintain or even exceed their original performance while retaining fewer than 5% of visual tokens, thereby reducing visual-token overhead and validating the necessity of dynamic token selection for practical clinical deployment. Our code is available at https://github.com/CUHK-AIM-Group/MedPruner.

Shengyuan Liu, Zanting Ye, Yunrui Lin +6

Chinese University of Hong Kong · Westlake University · Southern Medical University +4

Efficient VLM Inference 3D Medical Imaging Medical VLMs Visual Token Pruning

Ask PDF

May 1, 2026

Learning from Compressed CT: Feature Attention Style Transfer and Structured Factorized Projections for Resource-Efficient Medical Image Analysis

The deployment of artificial intelligence in medical imaging is hindered by high computational complexity and resource-intensive processing of volumetric data. Although chest computed tomography (CT) volumes offer richer diagnostic information than projection radiography, their use in AI-based diagnosis remains limited due to the computational burden of processing uncompressed volumetric images (typically stored in NIfTI or DICOM format). Addressing the growing need for low-resource deployment and efficient electronic data transfer, we investigate the utilization of JPEG-compressed chest CT volumes for thoracic abnormality detection. We propose Feature Attention Style Transfer (FAST), a novel distillation framework that transfers both activation patterns and structural relationships from high-fidelity CT representations to a spatiotemporal visual encoder operating on compressed inputs. By combining Gram-matrix-based attention style preservation with dual-attention feature alignment, FAST enables robust feature extraction from degraded volumes. Furthermore, we introduce Structured Factorized Projection (SFP), leveraging Block Tensor Train decomposition as a parameter-efficient alternative to dense projection layers, reducing projection-head parameters by almost half. Our contrastive learning pipeline, CT-Lite, integrates these components with a SigLIP-based multimodal alignment objective. Experiments on CT-RATE, NIDCH, and Rad-ChestCT demonstrate that CT-Lite achieves AUROC within 5-7% of the uncompressed-input baseline across all three datasets, despite operating on compressed inputs with significantly fewer parameters, paving the way for AI-based clinical evaluation under resource constraints.

Shadid Yousuf, S. M. Mahbubur Rahman, Mohammed Imamul Hassan Bhuiyan

Department of Electrical and Electronic Engineering, Bangladesh University of Engineering and Technology

Model Compression Medical Image Analysis Medical Image Anomaly Detection Computed Tomography Knowledge Distillation

Ask PDF

Jul 29, 2026

MedARC: Training-Free Adaptive Redundancy Compression of Visual Tokens for 3D Medical Vision-Language Models

Integrating 3D medical images with vision-language models (VLMs) holds substantial promise for computer-aided diagnosis. However, volumetric images generate prohibitively long visual-token sequences with considerable spatial and inter-slice redundancy. Existing token compression methods typically apply uniform reduction or rely on a single importance signal, increasing the risk of removing regions that are clinically relevant to the query or structurally distinctive. To address this limitation, we propose MedARC, a unified, training-free framework for Adaptive Redundancy Compression of visual tokens in 3D medical VLMs. MedARC estimates token importance by integrating three complementary cues: self-attention from the VLM vision encoder, which reflects the model's intrinsic visual focus; similarity between projected visual tokens and text embeddings, which identifies query-relevant regions; and deviations of local visual foundation model features from the volume-level feature center, which highlight structurally distinctive anatomy. The resulting importance distribution guides a saliency-aware merging strategy that preserves informative tokens while consolidating redundant ones rather than simply discarding them. Experiments on CT-RATE and MR-RATE show that MedARC reduces visual-token overhead and inference time while preserving or improving diagnostic performance. Its multi-cue scoring cost is outweighed by the savings from processing fewer tokens, with greater benefits expected for larger language models.

Yitao Zhu, Mengjun Liu, Yingji Fu +2

Department of Health Technology and Informatics, The Hong Kong Polytechnic University, Hong Kong · School of Integrated Circuits and Electronics, Beijing Institute of Technology, China · 3Mental Health Research Center, The Hong Kong Polytechnic University, Hong Kong +1

Efficient VLM Inference Medical VLMs Multimodal Token Compression Visual Token Compression 3D VLMs

Ask PDF