MedROI: Codec-Agnostic Region of Interest-Centric Compression for Medical Images

Jiwon Kim; Ikbeom Jang

cs.CVApr 6, 2026

MedROI: Codec-Agnostic Region of Interest-Centric Compression for Medical Images

Authors: Jiwon Kim, Ikbeom Jang

Organizations: Division of Computer Engineering, Hankuk University of Foreign Studies, Yongin, Republic of Korea · Department of Computer Science and Engineering, Hankuk University of Foreign Studies, Seoul, Republic of Korea

Abstract

Medical imaging archives are growing rapidly in both size and resolution, making efficient compression increasingly important for storage and data transfer. Most existing codecs compress full images/volumes(including non-diagnostic background) or apply differential ROI coding that still preserves background bits. We propose MedROI, a codec-agnostic, plug-and-play ROI-centric framework that discards background voxels prior to compression. MedROI extracts a tight tissue bounding box via lightweight intensity-based thresholding and stores a fixed 54byte meta data record to enable spatial restoration during decompression. The cropped ROI is then compressed using any existing 2D or 3D codec without architectural modifications or retraining. We evaluate MedROI on 200 T1-weighted brain MRI volumes from ADNI using 6 codec configurations spanning conventional codecs (JPEG2000 2D/3D, HEIF) and neural compressors (LIC_TCM, TCM+AuxT, BCM-Net, SirenMRI). MedROI yields statistically significant improvements in compression ratio and encoding/decoding time for most configurations (two-sided t-test with multiple-comparison correction), while maintaining comparable reconstruction quality when measured within the ROI; HEIF is the primary exception in compression-ratio gains. For example, on JPEG20002D (lv3), MedROI improves CR from 20.35 to 27.37 while reducing average compression time from 1.701s to 1.380s. Code is available at https://github.com/labhai/MedROI.

Figures & tables

Component	Fields	Type	Size (bytes)
Bounding Box	$\left(x_{\min},x_{\max},y_{\min},y_{\max},z_{\min},z_{\max}\right)$	int16	12
Original Shape	$(W,H,D)$	int16	6
Affine Matrix	$3\times 3$ rotation–scaling submatrix	float32	36
Total	–	–	54

Table 1: Metadata structure for ROI-based reconstruction.

Panel of Figure 1 — Figure 1: ROI-centric medical image compression. (a) 2D slice-wise compression pipeline: individual axial slices are extracted from the cropped volume and compressed/reconstructed independently. (b) 3D volumetric compression pipeline: ROI extraction, metadata generation, and volume-level codec compression/reconstruction. Both pipelines use metadata M (54 bytes) storing bounding box coordinates, original volume dimensions, and affine matrix for spatial restoration during decompression.

Model	Full/ROI	CR	BPP	PSNR	SSIM	Comp.(s)	Decomp.(s)
JPEG2000 lv3	Full	20.35	0.395	33.03	0.852	1.701	0.350
JPEG2000 lv3	ROI	27.37	0.397	30.73	0.859	1.380	0.265
JPEG2000 lv5	Full	5.165	1.567	45.14	0.969	1.722	0.734
JPEG2000 lv5	ROI	6.875	1.582	41.65	0.964	1.398	0.565
HEIF lv1	Full	39.38	0.241	29.38	0.842	0.059	1.5e-3
HEIF lv1	ROI	38.54	0.219	29.25	0.857	0.054	1.4e-3

Table 2: Average performance of 2D image compression methods with and without the MedROI approach. Compression and Decompression times are in seconds. comp.=compression time, decomp.=decompression time, CR=compression ratio, BPP=bits per pixel.

Model	Full/ROI	CR	BPP	PSNR	SSIM	Comp.(s)	Decomp.(s)
JPEG2000 lv1	Full	2.095	4.058	53.20	0.993	9.106	1.013
JPEG2000 lv1	ROI	2.316	3.621	52.38	0.994	6.729	0.233
JPEG2000 lv3	Full	4.758	1.755	47.06	0.981	5.898	1.023
JPEG2000 lv3	ROI	4.979	1.676	46.09	0.983	5.913	0.213
JPEG2000 lv5	Full	13.63	0.615	38.72	0.953	2.742	1.009
JPEG2000 lv5	ROI	14.23	0.591	37.49	0.949	2.446	0.204

Table 3: Average performance of 3D image compression methods evaluated with and without the MedROI approach. Total compression and decompression times are reported in seconds. comp.=compression time, decomp.=decompression time, CR=compression ratio, BPP=bits per pixel.

Figure 2: Rate-distortion performance of evaluated compression methods on the ADNI dataset. Each point represents the average PSNR and compression ratio across 200 brain MRI volumes. Solid markers indicate full-volume compression, while hollow markers represent ROI-based compression.

Figure 3: Qualitative comparison of Full vs ROI compression results. Visual comparison of reconstruction quality across different compression methods on subject 002_S_0413 (slice 128). Rows 1-2 show Full and ROI compression for 3D codecs (SirenMRI, JPEG2000-3D Lv3) and 2D conventional codec (JPEG2000-2D Lv3), while rows 3-4 display 2D neural compression codec (BCM-Net, TCM, TCM+AuxT) and HEIF Lv3. Ground truth is shown in the leftmost column, with the ROI bounding box (white rectangle) indicating the cropped region.

Explore similar work

MRIComp4Flow: Compression of 3D Brain MRI for Training Multi-Modal Generative Models
Aug 10, 2026Lisa K. Fischer, Mykhailo Riabets, Daniel Rueckert +33D Medical Imaging Medical Image Synthesis
Technical University of Munich (TUM), Germany · Chair for AI in Healthcare and Medicine, TUM and TUM University Hospital, Munich, Germany · Imperial College London +4
Hierarchical Spatio-Channel Clustering for Efficient Model Compression in Medical Image Analysis
Apr 25, 2026Sisipho Hamlomo, Marcellin Atemkeng, Habte Tadesse Likassa +5Low-Rank Compression Low-Rank Approximation
Department of Mathematics, Rhodes University, Makhanda 6140, South Africa · Department of Statistics, Rhodes University, Makhanda 6140, South Africa · National Institute for Theoretical and Computational Sciences, Stellenbosch 7600, South Africa +8
Practical Lossless Volumetric Medical Image Compression via Tri-plane Context Tree Learning
Aug 14, 2026Yuanchao Bai, Yifan Zhao, Kai Wang +53D Medical Imaging Learned Image Compression
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

Aug 10, 2026J/K move · Enter open · S save

MRIComp4Flow: Compression of 3D Brain MRI for Training Multi-Modal Generative Models

Lisa K. Fischer, Mykhailo Riabets, Daniel Rueckert, Benedikt Wiestler+2

Technical University of Munich (TUM), Germany · Chair for AI in Healthcare and Medicine, TUM and TUM University Hospital, Munich, Germany · Imperial College London+4

3D Medical Imaging Medical Image Synthesis

Open paper AskPDF arXiv

Large-scale multi-modal MRI datasets impose substantial storage and I/O costs, limiting the training of 3D generative models on commodity infrastructure. While lossy compression is known to preserve accuracy for discriminative segmentation networks, its effect on generative models, which must learn the full data distribution rather than a decision boundary, is unexplored. We study whether standard image codecs can effectively compress semantically rich brain tumor MRI while preserving the fidelity required to train and deploy a 3D MRI generative model. Each 3D volume is compressed with JPEG2000 or a near-lossless JPEG-LS pipeline. Next, a Wavelet Flow Matching model, conditioned on BraTS image sequences (T1n, T1c, T2, T2f), is trained on compressed data, and the resulting models are evaluated on the validation set. At a 20:1 compression ratio, synthesis quality is statistically equivalent to a model trained on uncompressed data within a pre-specified margin (

Δ

PSNR

<1

,dB,

Δ

SSIM

<0.02

; paired TOST

p=[[p]]

): mean PSNR is 27.3,dB vs. 27.0,dB and mean SSIM is 0.95 vs. 0.96 across modalities. Our results indicate that JPEG2000 compression is a practical step toward scalable 3D MRI generative modeling without degrading synthesis quality. The codebase is available at https://github.com/lisafis/MRIComp4Flow .

Abstract

Explore similar work

MRIComp4Flow: Compression of 3D Brain MRI for Training Multi-Modal Generative Models

Hierarchical Spatio-Channel Clustering for Efficient Model Compression in Medical Image Analysis

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