Dynamic-system neural framework compresses medical images up to 183-fold

Medical imaging is generating datasets that strain the storage and transmission capacity of clinical infrastructure. Photon-counting computed tomography, multiparametric MRI and whole-slide digital pathology sharpen diagnosis, but each scan adds to the data burden. A team now reports a compression framework called DSIN, short for dynamic system-empowered implicit neural representation, that achieves compression ratios of up to 183-fold while maintaining high reconstruction quality.

The work, published in Communications Engineering, addresses a trade-off in medical image compression. Current methods, the authors write, either sacrifice subvisual diagnostic features to prioritize human perception or lack generalizability across modalities and pathologies.

DSIN represents each image with an image-specific implicit neural representation, encoding the image as a learned function rather than a grid of pixel values. A stacked dynamic-system-based compression module further reduces the size of the model while preserving reconstruction quality, so the compressed output contains fewer parameters while still reproducing the original scan at high fidelity.

The researchers evaluated DSIN across five large-scale clinical datasets spanning computed tomography, MRI, positron emission tomography, computed tomography angiography and histopathology images. According to the paper, DSIN achieved compression ratios of up to 183-fold while maintaining high reconstruction quality across these datasets.

The team also examined how compressed images performed in downstream tasks such as segmentation and classification. DSIN outperformed both established compression methods and other implicit neural representation approaches on these benchmarks. That distinction matters for clinical workflows, since a compressed scan that degrades a segmentation model’s ability to trace an organ boundary could introduce errors even if the image looks acceptable to a radiologist.

Quantitative evaluations highlighted specific strengths: preserved reconstruction quality, maintained organ boundary fidelity and retained delineation of metabolic activity in low-dose PET imaging. Low-dose PET is sensitive territory for compression, because the signal revealing metabolic behavior is subtle and could be smoothed away by an aggressive algorithm. The authors state that DSIN provides an effective compression framework for the medical imaging modalities evaluated in the study.

The study’s scope is worth noting. The results cover the five modalities and datasets the team tested, and the authors frame DSIN as effective for the imaging modalities evaluated rather than as a universal solution. The article was released as an early shared version that is citable and carries a permanent DOI, with a final version of record to follow. Even so, the reported combination of high compression ratios, cross-modality coverage and preserved downstream task performance positions DSIN as a candidate for easing the data strain that accompanies ever more detailed medical imaging.

Subject of Research: A dynamic-system-based implicit neural representation framework for compressing multidimensional medical images across five clinical modalities.

Article Title: DSIN: dynamic system-empowered implicit neural representation for medical image compression

Article References: Song, H., Shi, H., Li, Y., Guo, H., Wang, H., Ma, J., Fan, F., & Zeng, D. (2026). DSIN: dynamic system-empowered implicit neural representation for medical image compression. Communications Engineering. https://doi.org/10.1038/s44172-026-00812-3

Image Credits: AI Generated

DOI: 10.1038/s44172-026-00812-3

Keywords: medical imaging, image compression, implicit neural representation, dynamic systems, computed tomography, MRI, PET, histopathology, deep learning, clinical infrastructure, segmentation, low-dose PET