From Sinograms to Self-Supervised Networks: Three Developmental Phases

Medical imaging noise reduction has moved through three distinct phases across the 70+ patent and literature records in this dataset, spanning 2000 to early 2026. Each phase reflects a generational shift in the underlying computational approach: from sinogram-domain preprocessing, through commercial iterative reconstruction products, to deep learning architectures now entering active patent protection.

The Foundational Phase (2000–2010) established the architectural vocabulary still visible in modern systems. Samsung Electronics’ 2000 US filing on low-power portable CT addressed noise via quantitative image restoration — the earliest patent in this dataset. The Research Foundation of State University of New York’s 2003 filing codified sinogram-domain followed by edge-preserving image-domain filtering as a two-stage approach. University of Rochester’s 2007 US patent — covering a pipeline combining digital reconstruction filter (DRF) and wavelet transform (WT) denoising embedded within the reconstruction process — established the multi-stage architecture that persists in commercial implementations filed across at least five jurisdictions.

The Algorithm Maturation Phase (2010–2020) was defined by OEM iterative reconstruction products: GE ASiR/ASiR-V, Siemens SAFIRE/ADMIRE, Philips iDose, and Canon AIDR3D. Noise power spectrum (NPS) methodology became the standard evaluation framework. A 2016 literature study documented GE’s Veo 3.0 model-based iterative reconstruction, while a 2013 cohort study demonstrated dose reductions without procedural quality degradation across 614 neuroangiography patients.

The Deep Learning Transition Phase (2020–2026) marks the period in which architectures including U-Net, residual dense networks, HRNet, GANs, and Transformer-based models moved from research to active patent filing. According to WIPO trend data, AI-related medical imaging patents have seen sustained multi-year growth — a pattern clearly visible within this dataset, where the majority of deep learning citations are concentrated in the most recent six years.

Four Technical Clusters Shaping the Patent Landscape

Medical imaging noise reduction patents and literature in this dataset organise into four distinct technical clusters, each reflecting a different computational philosophy and clinical deployment context. Understanding these clusters is essential for teams conducting freedom-to-operate analysis or technology gap assessments.

Cluster 1: Classical Spatial and Transform-Domain Filtering

Gaussian smoothing, median filtering, non-local means (NLM), block-matching 3D (BM3D), total variation (TV) minimisation, and wavelet thresholding dominated clinical practice through the 2010s. A 2021 phantom study demonstrated TV outperforming Wiener and median filters on PET/MR images. A 2020 systematic review synthesising 25 publications confirmed NLM as the leading non-deep-learning approach for brain MRI. Most strikingly for clinical ultrasound, a 2022 study reported NLM delivered a 28.62% CNR improvement for thyroid nodule imaging.

Wavelet-domain filtering has also generated commercial patent activity. Healcerion Co., Ltd.’s 2023 US patent integrates discrete wavelet transform (DWT)-based denoising into a remote medical diagnosis platform supporting X-ray, ultrasound, CT, and MRI — a commercially oriented extension of a foundational algorithm class into connected-health infrastructure.

Cluster 2: Iterative Reconstruction and Model-Based Algorithms

Iterative reconstruction replaces or supplements filtered back projection with forward model-based or statistical optimisation frameworks. University of Rochester’s multi-jurisdiction patent family — covering global cone beam CT denoising by combining DRF and wavelet transforms at multiple pipeline stages — is among the most widely filed foundational contributions in this dataset, spanning at least five jurisdictions (WO, EP, US, CA, AU, IN). Philips’ 2025 JP filing extends the iterative tradition into spectral CT by applying a neural network trained for the dual task of material decomposition and denoising in dual-energy CT projection data.

Cluster 3: Deep Learning and Neural Network Architectures

This is the fastest-growing cluster, with the majority of deep learning citations concentrated in 2020–2026. A 2021 study introducing HRNet (High-Resolution Network) for low-dose CT denoising demonstrated superiority over U-Net baselines. A 2017 paper pioneered applying CNN denoising to wavelet transform coefficients of CT images. GE Precision Healthcare holds two active US patents — filed in 2021 and updated in 2024 — applying deep learning in commercial CT settings, signalling an active continuation filing strategy around deployed products.

Cluster 4: Modality-Specific Denoising (OCT, MRI, PET, Fluoroscopy)

Several patent families address noise characteristics unique to individual modalities. For OCT, Korea University Research and Business Foundation holds two active US patents (2013, 2016) on adjacent-depth-data subtraction for frequency-domain OCT. Topcon Corporation’s 2026 US patent specifies machine learning operation at 400 kHz A-line rates, with separate ML models for distinct noise types. For low-field MRI, deep learning reconstruction using the AUTOMAP end-to-end network reported SNR gains of 1.5 to 4.5-fold at 6.5 mT field strength. For fluoroscopy, Philips AlluraClarity NRT demonstrated a 35% reduction in cumulative air kerma during adrenal vein sampling.

“Topcon’s 2026 OCT patent specifies operation at 400 kHz A-line rates with separate ML models for distinct noise types — real-time, noise-type-aware denoising is moving from research prototype to commercial product readiness.”

Modality Application Domains: Where the Clinical Evidence Is Strongest

CT accounts for the largest concentration of records in this dataset, driven by radiation dose reduction imperatives across head CT, abdominal CT, lung cancer screening via low-dose CT (LDCT), cardiac imaging, and cone-beam CT. Iterative reconstruction allowed head CT dose reductions of 30–60% without compromising diagnostic quality, according to a 256-slice MDCT study. Lung cancer LDCT screening is addressed by a 2021 study applying genetic algorithm-optimised wavelet thresholding — connecting classical transform methods to modern optimisation approaches.

Interventional radiology and fluoroscopy represent the highest-stakes radiation reduction context, where both operator and patient exposure accumulate across lengthy procedures. Philips AlluraClarity NRT is documented in TAVR and structural heart procedures (2018 conference abstract) and in a 614-patient neuroangiography cohort study (2013) demonstrating dose reductions without procedural quality degradation. These large cohort data points are among the strongest clinical evidence in this dataset for quantifiable dose reduction at scale.

MRI noise reduction spans both high-field (3T, 7T) and emerging low-field systems. GE Healthcare’s AIR Recon DL for prostate MRI at 1.5T/3T was commercially available as of the 2021 evaluation study. At the research frontier, a 2023 study evaluating deep learning reconstruction for preclinical 7T MRI reported SNR improvement from 10.4 to 32.1 in rat brain cortex using residual dense networks, demonstrating the quality ceiling that deep learning can now approach in high-field preclinical settings. This body of work is tracked by bodies including NIH’s National Institute of Biomedical Imaging and Bioengineering, which funds MRI reconstruction research across field strengths.

OCT noise reduction addresses speckle and fixed-pattern noise across ophthalmology, dermatology, and cardiology applications. The 2010 curvelet transform speckle suppression study and Topcon’s 2026 machine learning patent bracket a decade of OCT-specific innovation — from mathematical transform approaches to real-time ML models operating at 400 kHz A-line rates. Nuclear medicine applications, including PET/MR and gamma camera SPECT, are addressed through total variation algorithms and median filter variants documented in 2019 and 2021 literature studies. Standards for evaluating these systems are developed in part by organisations including IAEA and NEMA.

Geographic and Assignee Concentration Patterns

Innovation in medical imaging noise reduction, as captured in this dataset, is concentrated in a relatively small number of institutional filers, with a longer tail of academic and startup contributors. The United States dominates as the primary jurisdiction for active patent filings, with China showing acceleration in recent years.

Among US-based or US-filing assignees: Mayo Foundation for Medical Education and Research holds the most recent and strategically significant filings, with a 2024 WO patent on deep learning for VOI CT and a 2025 US patent on uncertainty quantification. GE Precision Healthcare LLC holds active US patents from 2021 and 2024. University of Rochester’s foundational pipeline patent spans at least five jurisdictions. ALARA Systems, Cornell University, Carl Zeiss X-Ray Microscopy Inc., and NIDEK Co., Ltd. round out the US-oriented roster.

Chinese filers are accelerating in low-dose DR and CT denoising. Guangzhou BaiShi Medical Technology Co., Ltd. holds two active CN patents (2020, 2023) combining adaptive Gaussian frequency-domain filtering with NLM and deep neural networks. Xi’an University of Technology’s 2025 CN patent deploys a dual-channel Transformer-based architecture for low-dose CT. Sun Yat-sen University filed a low-dose DR/CT denoising patent in 2023. Product developers entering or competing in the Chinese market must account for this growing domestic IP position, as noted in WIPO‘s annual IP statistics reporting China’s sustained growth in medical device patent filings.

South Korean assignees are concentrated in OCT: Korea University Research and Business Foundation holds two active US patents (2013, 2016), and Healcerion Co., Ltd. holds two active US patents in DWT-based noise removal for remote medical diagnosis. Japanese assignees include Topcon Corporation (2026 US filing) and NIDEK Co., Ltd. (US and EP patents in OCT tomographic noise reduction). European filings include University of Rochester (EP 2008), Arjae Spectral Enterprises (EP 2010), and Carl Zeiss X-Ray Microscopy (EP 2025).

Six Emerging Directions in 2024–2026 Filings

The most recent filings in this dataset signal six convergent directions that characterise the leading edge of medical imaging noise reduction innovation — each representing a shift from general-purpose denoising toward task-specific, modality-tuned, and clinically actionable AI architectures.

1. Deep Learning for VOI and Artifact-Specific Denoising. Mayo Foundation’s 2024 WO patent targets truncation artifact removal in helical volume-of-interest (VOI) CT — a highly specific application that conventional filters cannot address. This signals a move from general-purpose denoising toward task-specific deep learning modules embedded in acquisition pipelines.
2. Uncertainty Quantification for Clinical Decision Support. Mayo Foundation’s 2025 US patent introduces pixel-wise uncertainty maps derived from bootstrap approximation or trained ML algorithms for CNN-based noise reduction. This provides radiologists with spatial confidence maps and is a critical step toward regulatory acceptance of AI-based denoising, as regulatory bodies — including the FDA — are increasingly requiring explainability outputs for cleared AI/ML imaging devices.
3. Super-Resolution Combined with Noise Reduction. Carl Zeiss X-Ray Microscopy’s dual US and EP filings in December 2025 integrate noise suppression into a multiscale super-resolution framework for X-ray nanotomography, pointing toward joint denoising-and-resolution-enhancement as a unified computational task.
4. High-Speed OCT with Machine Learning at the Sensor Layer. Topcon’s 2026 US patent specifies operation at 400 kHz A-line rates with separate ML models for distinct noise types — suggesting real-time, noise-type-aware denoising is moving from research prototype to commercial product readiness.
5. Dual-Energy CT Denoising via Unified Neural Decomposition. Philips’ 2025 JP filing combines material decomposition and denoising into a single neural network trained end-to-end, replacing the sequential two-step pipeline that has historically dominated spectral CT workflows.
6. Dual-Channel Transformer-Based CT Denoising. Xi’an University of Technology’s 2025 CN patent deploys parallel U-Net branches — one based on sparse Transformer blocks, one on residual attention modules — with output fusion, reflecting adoption of attention mechanisms in Chinese academic and commercial filings.

“Mayo Foundation’s 2025 patent on pixel-wise uncertainty maps for CNN noise reduction represents a nascent but strategically important area — one that is underprotected relative to its likely regulatory significance.”

Strategic Implications for IP and R&D Teams

The patent and literature signals in this dataset translate into five actionable strategic observations for IP counsels, R&D directors, and technology strategists operating in the medical imaging space.

Deep learning denoising is transitioning from research publication to active patent protection. R&D teams should monitor continuation filings from Mayo Foundation, GE Precision Healthcare, Siemens Healthineers, Philips, and Topcon. These assignees are actively building IP positions around commercially deployed AI reconstruction systems, not merely exploratory algorithms.

The uncertainty quantification vector is underprotected. Mayo Foundation’s 2025 filing on uncertainty maps for CNN noise reduction represents a nascent but strategically important area. IP strategists for OEMs and clinical AI companies should evaluate freedom-to-operate and differentiation opportunities around confidence visualisation, as regulatory bodies are likely to require such outputs for cleared AI/ML-based imaging devices.

China-based assignees are accelerating in DR and CT denoising. Guangzhou BaiShi Medical Technology Co., Ltd. and Xi’an University of Technology hold active CN patents combining classical filters with deep networks for low-dose DR and CT. Product developers entering the Chinese market or competing with Chinese OEMs must account for this growing domestic IP position.

Joint super-resolution and denoising architectures represent a white-space opportunity. Carl Zeiss’s 2025 filings combine these objectives in the X-ray nanotomography domain. The equivalent approach for clinical CT or MRI remains lightly patented in this dataset, suggesting an open area for differentiated IP development.

Modality-specific noise characteristics require tailored solutions. The dataset confirms that OCT speckle, MRI k-space noise, PET Poisson statistics, and CT quantum noise each demand distinct algorithmic treatments. Platform strategies that attempt one-size-fits-all denoising modules will face performance limitations. Modality-tuned, task-specific architectures — as seen in Topcon’s 2026 OCT patent with separate ML models for distinct noise types — represent the emerging commercial standard. This specialisation trend aligns with recommendations from standards bodies such as IAEA on nuclear medicine image quality and is reflected in evaluation frameworks published by NEMA.