West China Hospital Licenses Breakthrough CT Metal Artifact Reduction Technology for Over RMB 200,000

Recently, West China Hospital of Sichuan University released a public notice proposing to“A Method for Metal Artifact Removal in CT Images Based on a Dual-Domain Progressive Enhancement Network”Invention patent, licensed to industry partners for commercialization through a non-exclusive license. The transaction price is determined by agreement, with the proposed price beingRMB 200,000RMB fixed fee, plusSales Commission (RMB 2,500 per unit; the first four units are excluded from commission)。

This patented technology was jointly developed by inventors including Wu Min, Zhao Qijun, Tan Jia, Xiong Deng, and Yao Xiaoli from West China Hospital, aiming to address the severe artifacts caused by metal implants in CT images. Its core design establishes aIntelligent Closed-Loop Processing System, innovatively integrating a dual-domain progressive enhancement network with an embedded quantitative assessment mechanism.

This system not only effectively removes artifacts and preserves anatomical details through the synergistic operation of networks in the sinogram and image domains, but also automatically performs quality assessment and decision-making during critical stages such as data processing and model training. This systematically enhances the precision and reliability of artifact removal, thereby providing clearer diagnostic images for clinical practice.

The advancement of technology translation stems from clear challenges that have long existed in the clinical practice of medical imaging. Metal artifacts have always been a "stubborn disease" in CT scans. With the development of modern medical technology, it is becoming increasingly common for patients to have various metal implants in their bodies, ranging from artificial joints and spinal internal fixation devices in orthopedics, to dental fillings and implants, as well as cardiovascular stents.

These high-density materials cause severe streak and starburst artifacts in CT imaging, significantly interfering with the visualization of surrounding normal tissues, particularly fine anatomical structures, and directly impacting clinicians’ diagnosis of lesions, surgical planning, and assessment of treatment efficacy.

Traditional solutions often face a dilemma.

On one hand, methods based on classical image processing algorithms, such as filtering or interpolation, often result in image blurring or introduce new artifacts while removing existing ones, leading to a significant loss of valuable diagnostic details. On the other hand, although deep learning methods that have emerged in recent years show promise, their model training heavily relies on large-scale, paired (i.e., corresponding artifact-laden and artifact-free) high-quality annotated data.

In practice, acquiring such data is costly and labor-intensive, requiring specialized medical expertise for precise annotation. This leads to data scarcity, which becomes a bottleneck that limits the model's generalization capability and robustness in real-world complex scenarios.

Therefore, there is an urgent need in the medical field for a technical solution that can efficiently remove artifacts while maximizing the preservation of critical diagnostic information, and that relies on training data in a more intelligent and robust manner. The research and development team at West China Hospital is focusing precisely on this core pain point driven by clinical needs.

In response to the aforementioned clinical needs and technical bottlenecks, the team did not limit itself to optimizing a single algorithm, but instead constructedA Systematic, Closed-Loop Driven Intelligent SolutionIts core value lies in upgrading metal artifact removal from a post-processing step with “uncertain outcomes” to a reliable technology with “controllable processes and assessable quality,” through tightly coupled innovations across three levels.

· A three-tier processing architecture featuring "division of labor with collaboration and progressive enhancement."The patented network model follows a clear logical chain:

First, at the raw data level (sinogram domain), throughAttention Embedding MechanismPrecisely locate and correct projection data severely contaminated by metal artifacts, striving to minimize artifact introduction at the source. Subsequently, the team performed operations in the image domain of the preliminary reconstructionSecondary Refined Repair, eliminating residual noise and artifacts. Finally, the team specifically introducedIndependent Detail Enhancement Network, with its core task being to restore and enhance anatomical edges and tissue textures that may have been blurred during the previous two processing steps.

This“Projection Domain Restoration → Image Domain Refinement → Global Detail Recovery”process, ensuring that artifact removal and detail preservation are no longer a trade-off but objectives that can be achieved simultaneously.

· Also the most groundbreaking is the “Embedded Quality Assessment and Decision-Making Hub” that spans the entire process.This is precisely the key feature that distinguishes this technology from traditional “black box” models. It establishes intelligent checkpoints at three critical nodes:

Incoming Quality Inspection Station:Prior to data ingestion into the model, the system automatically evaluates the quality of preprocessed data using an Image Processing Quality Assessment Index. Only data meeting the established thresholds are approved for training, thereby ensuring robust model performance from the outset.

Output Verification Station:Upon completion of model training, the system conducts a multi-dimensional quantitative assessment of its performance on the test set (using an artifact removal compliance evaluation index), covering multiple clinically relevant dimensions such as structure, edges, and texture. Failure to meet the predefined standards automatically triggers an optimization process.

Optimization Feedback Station:Any optimization applied to the model is quantitatively assessed using an Optimization Effectiveness Evaluation Index to determine its efficacy and inform decisions on whether to revert to earlier stages (e.g., data re-preparation). This mechanism empowers the systemSelf-Diagnosis, Self-Optimizationthe capability, significantly reducing reliance on “perfect training data” and enhancing the technology’s performance in complex real-world scenariosAdaptability and Robustness.

· It is the “quantitative evaluation system” for achieving decision intelligence.All the aforementioned assessments and decisions are not based on subjective experience, but rather rely on a series of clear, computable mathematical metrics. The multiple assessment indices defined in the patent transform the abstract concept of image quality into objective numerical values by integrating multi-dimensional parameters such as noise level, structural similarity, and processing efficiency. This renders the operational status and processing efficacy of the entire systemTransparent, Measurable, Traceable。

In summary, the innovative value of this patent far exceeds that of a merely superior neural network. It provides a comprehensive methodology and engineering framework encompassing data quality control, intelligent processing, and efficacy validation.Its translation not only signifies the implementation of an advanced algorithm, but also represents a kind of# Pursuing Certainty and High-Quality Outcomestechnical paradigm, is expected to provide more reliable tools for the precise diagnosis of CT images, with clear clinical significance and market prospects.

This patented technology from West China Hospital emerged in a field characterized by intense competition and rapid iteration among research institutions and enterprises, both domestically and internationally. A comparative analysis more clearly highlights its unique value.

Domestic technological development is vibrant, with dual-domain and weakly supervised learning emerging as hot topics.In recent years, domestic universities, hospitals, and technology companies have made significant progress in this field. For instance, a patent granted in 2025 (CN120563661B) proposed a “CBCT Metal Artifact Suppression Method Based on Dual-Domain Multi-View Weakly Supervised Segmentation,” with its innovation lying in the use of multi-view consistent segmentation and weakly supervised learning to reduce reliance on precisely annotated data. Meanwhile, the industry is also actively strategizing; for example, Tencent Corporation published a patent in 2025 (CN120355800A) that removes artifacts by simulating the metal diffusion process.

These developments collectively reflect two core directions of current technological exploration in China: first, deepeningDual-Domain Information Fusion, secondly, to seekReducing Reliance on Paired Annotation Data。

International research frontiers focus on unsupervised learning and physics-informed approaches.Globally, in addition to major medical device manufacturers (such as GE and Siemens) integrating advanced iterative reconstruction and dual-energy spectral artifact reduction technologies into their high-end CT systems, the academic research frontier is exhibiting new trends. On one hand,Unsupervised or Weakly Supervised LearningHighly valued for its ability to completely eliminate reliance on hard-to-acquire paired data. For example, a study in 2025 proposed an unsupervised deep sparse transformation network to jointly achieve metal artifact reduction and super-resolution reconstruction.

On the other hand, exploreDeep Integration with Physical Imaging Models or Prior Knowledgekey to enhancing the robustness and interpretability of enhancement methods. For instance, studies have attempted to align data in the latent space using Gemstone Spectral Imaging to more accurately suppress artifacts. These efforts indicate that the international frontier is dedicated to developing more generalizable, robust, and physics-compliant intelligent artifact removal models.

In this technology landscape, West China Hospital’s patents exhibit a distinct characteristic of systematic engineering thinking. Unlike many studies that focus on improving network modules or loss functions, the core competitiveness of this technology lies in its construction of“Embedded Assessment and Closed-Loop Decision-Making”Framework. It does not merely pursue higher Peak Signal-to-Noise Ratio (PSNR) on specific datasets, but instead incorporates the Image Quality assessment index (IQ), the Artifact Removal compliance assessment index (AR), and the Optimization Effectiveness assessment index (EA), inPreprocessing, Model Training, Result Output, Iterative OptimizationQuality monitoring and automated decision-making nodes are embedded throughout the entire process.

This enables the technical solution to possess greater robustness when confronted with the complex and highly variable real-world data encountered in clinical practice.Adaptive CapacityandControllability of OutcomesIn short, its innovation lies not only in “how to better remove artifacts” but also in “how to systematically ensure that each removal is reliable and usable,” thereby providing the critical engineering safeguards needed for its transition from the laboratory to real-world clinical applications.

* Patent transaction information provided by CSTT