A frequency-adaptive implicit neural representation method for medical image compression

doi:10.12122/j.issn.1673-4254.2026.07.23

Abstract

Abstract:

Objective To address the limitations of data storage and transfer caused by exponential growth of medical imaging data size, we propose a frequency-adaptive implicit neural compression (FAINC) method for medical images based on optimized implicit neural networks (INRs). Methods A retrospective analysis was conducted on abdominal CT data from 356 patients in the KiTS19 and AVT datasets. We developed the FAINC method, a multi-subnetwork collaborative compression framework, which first evaluates the frequency-domain complexity of image blocks using the Spectral Sparsity Index (SSI), and then dynamically allocates them to subnetworks of different capacities through a frequency-domain gating mechanism. Compression output is achieved by combining parameter quantization with entropy coding. To assess its performance, the proposed method was compared with mainstream commercial compression standards (H.265/HEVC and JPEG2000), the implicit neural representation method NeRV, and the deep-learning-based compression method DVC. Results The FAINC method achieved the best reconstruction performance on both KiTS19 and AVT datasets at high compression ratios. At bitrates of BPV=0.32 and BPV=0.34, the FAINC method obtained the highest PSNR (47.03 and 50.76), the highest SSIM (0.9853 and 0.9930), and the lowest RMSE (0.0045 and 0.0029), achieving also a significantly higher subjective image quality score than other methods. Ablation studies demonstrated that the frequency-domain gating mechanism and dynamic parameter allocation contributed approximately 2.05 dB and 1.87 dB PSNR improvements, respectively. Conclusion The proposed method substantially enhances image reconstruction quality at high compression ratios and outperforms the existing mainstream approaches in terms of structural fidelity and compression efficiency. The FAINC method provides a promising technical solution for efficient storage and low-bandwidth remote transfer of medical image data.

Key words: medical image compression, implicit neural representation, frequency-adaptive modeling, high-fidelity reconstruction

Hanxiao SONG, Huaxian SHI, Ying LI, Zhaoying BIAN, Dong ZENG. A frequency-adaptive implicit neural representation method for medical image compression[J]. Journal of Southern Medical University, 2026, 46(7): 1703-1713.

Figures/Tables 12

Fig.1 Overall architecture of the proposed frequency-adaptive implicit neural compression (FAINC) method.

Fig.2 Rate distortion curves of different compression methods on the KiTS19 dataset. A: BPV-PSNR curves. B: BPV-SSIM curves.

Tab.1 Mean and variance of reconstruction metrics for different methods on 20 3D CT volumes from the KiTS19 dataset with BPV of 0.32-0.34

Methods	PSNR	PSNR (Rel. ↑ %)	SSIM	SSIM (Rel. ↑ %)	RMSE	RMSE (Rel. ↓ %)
FAINC	47.0346±0.7289	-	0.9853±0.0015	-	0.0045±0.0004	-
JPEG2000	45.7261±0.4892	2.86	0.9800±0.0016	0.54	0.0052±0.0003	13.46
H.265	43.8230±0.2628	7.34	0.9629±0.0039	2.32	0.0064±0.0002	29.69
H.264	42.0655±0.3790	11.81	0.9539±0.0038	3.30	0.0079±0.0004	43.04
JPEG	42.0379±0.3820	11.87	0.9539±0.0039	3.30	0.0079±0.0004	43.04
NeRV	43.7035±1.2676	7.62	0.9716±0.0116	1.41	0.0124±0.0016	63.71

Fig.3 Visual comparison of reconstructed images produced by different methods on the KiTS19 dataset under similar compression ratios (BPV=0.34-0.36). A1-A3: Ground Truth; B1-B3: FAINC; C1-C3: H.265; D1-D3: H.264; E1-E3: NeRV; F1-F3: JPEG. The display window is [-160, 240]HU.

Tab.2 Overall image quality scores for reconstructed images after FAINC compression of the KiTS19 dataset

Method	Mean±SD	P
FAINC	4.72±0.45
JPEG2000	4.62±0.58	0.019
H.265	4.23±0.70	<0.001
H.264	3.97±0.64	<0.001
JPEG	3.45±0.75	<0.001
NeRV	2.35±0.61	<0.001

Fig.4 Rate distortion curves of different compression methods on the AVT dataset. A: BPV-PSNR curves. B: BPV-SSIM curves.

Tab.3 Mean and variance of reconstruction metrics for different methods on 18 3D CT volumes from the AVT dataset (BPV=0.33-0.35)

Methods	PSNR	PSNR (Rel. ↑ %)	SSIM	SSIM (Rel. ↑ %)	RMSE	RMSE (Rel. ↓ %)
FAINC	50.7630±0.6820	-	0.9930±0.0010	-	0.0029±0.0004	-
JPEG2000	50.9773±1.2549	-0.42	0.9918±0.0019	0.12	0.0031±0.0004	6.45
H.265	49.1827±0.9431	3.21	0.9872±0.0085	0.59	0.0035±0.0004	17.14
H.264	44.8626±1.0237	13.15	0.9564±0.0117	3.83	0.0058±0.0008	50.00
JPEG	44.8287±1.0157	13.24	0.9563±0.0117	3.84	0.0058±0.0008	50.00
NeRV	43.1059±4.5098	17.77	0.9701±0.0220	2.36	0.0047±0.0014	38.30

Fig.5 Visual comparison of reconstructed images produced by different methods on the AVT dataset under similar compression ratios (BPV=0.34-0.36). A1-A3: Ground Truth; B1-B3: FAINC; C1-C3: H.265; D1-D3: H.264; E1-E3: NeRV; F1-F3: JPEG. The display window is [-160, 240]HU.

Tab.4 Overall image quality scores for reconstructed images after FAINC compression of the AVT dataset

Method	Mean ± SD	P
FAINC	4.77±0.43
JPEG2000	4.55±0.50	0.005
H.265	4.20±0.58	0.002
H.264	3.92±0.46	<0.001
JPEG	3.55±0.65	<0.001
NeRV	2.72±0.69	<0.001

Fig.6 Visual comparison of reconstructed medical images by FAINC and two ablation models. A1-A2: Ground Truth; B1-B2: FAINC; C1-C2: W/o Frequency-Guided Partition; D1-D2: W/o Adaptive Allocation. The display window is [-160, 240]HU.

Fig.7 Comparison of rate-distortion curves for FAINC and two ablation models.

Fig.8 PSNR and parameter scale comparison across multiple subnetworks for FAINC and two ablation models. Top: Reconstruction quality (PSNR); bottom: Sub-network memory usage (MB).

References 34

[1]	Minopoulou I, Kleyer A, Yalcin-Mutlu M, et al. Imaging in inflammatory arthritis: progress towards precision medicine[J]. Nat Rev Rheumatol, 2023, 19(10): 650-65. doi：10.1038/s41584-023-01016-1
[2]	Moore J, Allan C, Besson S, et al. OME-NGFF: a next-generation file format for expanding bioimaging data-access strategies[J]. Nat Methods, 2021, 18(12): 1496-8. doi：10.1038/s41592-021-01326-w
[3]	Wiseman S. A FAIR platform for data-sharing[J]. Nat Neurosci, 2021, 24(12): 1640. doi：10.1038/s41593-021-00976-5
[4]	Boergens KM, Berning M, Bocklisch T, et al. webKnossos: efficient online 3D data annotation for connectomics[J]. Nat Methods, 2017, 14(7): 691-4. doi：10.1038/nmeth.4331
[5]	Wallace GK. The JPEG still picture compression standard[J]. Commun ACM, 1991, 34(4): 30-44. doi：10.1145/103085.103089
[6]	Wiegand T, Sullivan GJ, Bjontegaard G, et al. Overview of the H.264/AVC video coding standard[J]. IEEE Trans Circuits Syst Video Technol, 2003, 13(7): 560-76. doi：10.1109/tcsvt.2003.815165
[7]	Wien M. High efficiency video coding[J]. Coding Tools and specification, 2015, 24: 1. doi：10.1007/978-3-662-44276-0
[8]	Dong C, Deng YB, Loy CC, et al. Compression artifacts reduction by a deep convolutional network[C]//2015 IEEE International Conference on Computer Vision (ICCV). December 7-13, 2015. Santiago, Chile. IEEE, 2015: 576-84. doi：10.1109/iccv.2015.73
[9]	Rufai AM, Anbarjafari G, Demirel H. Lossy medical image compression using Huffman coding and singular value deco-mposition[C]//2013 21st Signal Processing and Communications Applications Conference (SIU). April 24-26, 2013, Haspolat, Turkey. IEEE, 2013: 1-4. doi：10.1109/siu.2013.6531592
[10]	Song CY, Hui C, Lin Q, et al. LVPNet: a latent-variable-based prediction-driven end-to-end framework for lossless compression of medical images[C]//2025 Medical Image Computing and Computer Assisted Intervention(MICCAI). Sept 23-27, 2025, Daejeon, Republic of Korea. Springer Nature Switzerland, 2025: 288-98. doi：10.1007/978-3-032-04984-1_28
[11]	Duan XT, Liu JJ, Zhang E. Efficient image encryption and compression based on a VAE generative model[J]. J Real Time Image Process, 2019, 16(3): 765-73. doi：10.1007/s11554-018-0826-4
[12]	Sitzmann V, Martel JNP, Bergman AW, et al. Implicit neural representations with periodic activation functions[C]//2020 Neural Information Processing Systems Conference (NeurIPS). Dec 6-12, 2020, San Diego, USA. Curran Associates, Inc, 2020: 7462-7473. doi：10.48550/arXiv.2006.09661
[13]	Dai GL, Zhang RY, Wuwu Q, et al. Implicit neural image field for biological microscopy image compression[J]. Nat Comput Sci, 2025, 5(11): 1041-50. doi：10.1038/s43588-025-00889-4
[14]	Mildenhall B, Srinivasan PP, Tancik M, et al. NeRF: representing scenes as neural radiance fields for view synthesis[J]. Commun ACM, 2022, 65(1): 99-106. doi：10.1145/3503250
[15]	Dupont E, Goliński A, Alizadeh M, et al. Coin: Compression with implicit neural representations[EB/OL]. 2021: arXiv:2103.03123. . doi：10.48550/arXiv.2103.03123
[16]	Chen H, He B, Wang H, et al. Nerv: Neural representations for videos［C］//2021 Neural Information Processing Systems Conference (NeurIPS). Dec 6-14, 2021, San Diego, USA. Curran Associates, Inc, 2021: 21557-68. doi：10.48550/arXiv.2110.13903
[17]	Lu Y, Jiang K, Levine JA, et al. Compressive neural representations of volumetric scalar fields[J]. Comput Graph Forum, 2021, 40(3): 135-46. doi：10.1111/cgf.14295
[18]	Bird T, Ballé J, Singh S, et al. 3D scene compression through entropy penalized neural representation functions[C]//2021 Picture Coding Symposium (PCS). June 29-July 2, 2021, Bristol, United Kingdom. IEEE, 2021: 1-5. doi：10.1109/PCS50896.2021.9477505
[19]	Yang RZ, Xiao TX, Cheng YX, et al. Sharing massive biomedical data at magnitudes lower bandwidth using implicit neural function[J]. Proc Natl Acad Sci USA, 2024, 121(28): e2320870121. doi：10.1073/pnas.2320870121
[20]	Ma Y, Yi C, Zhou Y, et al. Semantic redundancy-aware implicit neural compression for multidimensional biomedical image data[J]. Communications Biology, 2024, 7(1): 1081. doi：10.1038/s42003-024-06788-0
[21]	Yang RZ, Xiao TX, Cheng YX, et al. SCI: a spectrum concentrated implicit neural compression for biomedical data[J]. Proc AAAI Conf Artif Intell, 2023, 37(4): 4774-82. doi：10.1609/aaai.v37i4.25602
[22]	Dupont E, Loya H, Alizadeh M, et al. Coin++: Neural compression across modalities[EB/OL]. 2022: arXiv: 2201.12904. . doi：10.48550/arXiv.2201.12904
[23]	Sheibanifard A, Yu HC. A novel implicit neural representation for volume data[J]. Appl Sci, 2023, 13(5): 3242. doi：10.3390/app13053242
[24]	Liang R, Sun H, Vijaykumar N. Coordx: Accelerating implicit neural representation with a split mlp architecture[EB/OL]. 2022: arXiv:2201.12425. . doi：10.48550/arXiv.2201.12425
[25]	Xu JS, Moyer D, Gagoski B, et al. NeSVoR: implicit neural representation for slice-to-volume reconstruction in MRI[J]. IEEE Trans Med Imag, 2023, 42(6): 1707-19. doi：10.1109/tmi.2023.3236216
[26]	Jacobs RA, Jordan MI, Nowlan SJ, et al. Adaptive mixtures of local experts[J]. Neural Comput, 1991, 3(1): 79-87. doi：10.1162/neco.1991.3.1.79
[27]	Heller N, Sathianathen N, Kalapara A, et al. The KiTS19 challenge data: 300 kidney tumor cases with clinical context, CT semantic segmentations, and surgical outcomes[EB/OL]. 2019: arXiv: 1904.00445. . doi：10.48550/arXiv.1904.00445
[28]	Radl L, Jin Y, Pepe A, et al. AVT: Multicenter aortic vessel tree CTA dataset collection with ground truth segmentation masks[J]. Data Brief, 2022, 40: 107801. doi：10.1016/j.dib.2022.107801
[29]	Gong YC, Liu L, Yang M, et al. Compressing deep convolutional networks using vector quantization[EB/OL]. 2014: arXiv: 1412.6115. .
[30]	Moffat A. Huffman coding[J]. ACM Comput Surv, 2020, 52(4): 1-35. doi：10.1145/3342555
[31]	Skodras A, Christopoulos C, Ebrahimi T. The JPEG 2000 still image compression standard[J]. IEEE Signal Process Mag, 2001, 18(5): 36-58. doi：10.1109/79.952804
[32]	SJLjTomar. Converting video formats with FFmpeg[J]. Linux journal, 2006, 2006(146):10.
[33]	Lu G, Ouyang WL, Xu D, et al. DVC: an end-to-end deep video compression framework[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). June 15-20, 2019. Long Beach, CA, USA. IEEE, 2019: 10998-1007. doi：10.1109/cvpr.2019.01126
[34]	McDole K, Guignard L, Amat F, et al. In toto imaging and reconstruction of post-implantation mouse development at the single-cell level[J]. Cell, 2018, 175(3): 859-76. e33. doi：10.1016/j.cell.2018.09.031