Description:A METHOD AND SYSTEM FOR IMAGING DATASET GENERATION IMITATING CORONARY ANGIOGRAPHIC CONDITIONS

FIELD OF INVENTION
The present invention relates to the field of medical image processing. More specifically, it concerns a method and system for real-like imaging dataset generation that imitates coronary angiographic conditions enabling the development, training and validation of machine learning models. These models are designed to enhance coronary angiography image quality by effectively reducing noise, thereby improving the clarity, interpretability, and diagnostic accuracy of coronary vessel visualization.

BACKGROUND OF THE INVENTION
Coronary angiography is performed by introducing contrast dye into the coronary arteries and capturing X‑ray images to visualize vessel structure and assess the presence of blockages. Traditionally, this imaging method has been regarded as the gold standard for diagnosing coronary artery disease, offering clinicians a direct view of arterial narrowing or obstruction. However, the quality of angiographic images is frequently affected by various forms of noise, including motion artifacts, low‑dose radiation interference, and system‑related distortions. These issues reduce image clarity, making it difficult for clinicians to observe fine vessel structures and complicating accurate, particularly in intricate or small vessel regions. This limitation compromises diagnostic accuracy and treatment planning, leading to suboptimal patient outcomes.
Existing challenges in coronary angiography include the persistent issue of noise artifacts. Angiographic images frequently contain various types of noise, including temporal noise and geometric noise that arise from imaging techniques, patient movement, and low dose radiation settings. Noise can obscure important vessel details, making it difficult to assess vessel diameter, plaque build-up, and other markers essential for diagnosis. These imaging degradations represent a daily challenge in clinical catheterization laboratories.
Traditional noise reduction methods, such as Gaussian smoothing and median filtering, have clear limitations. While they may reduce some visible noise, they often blur vessel edges and erase subtle but important anatomical detail. As a result, these approaches degrade the structural integrity of the images and reduce diagnostic confidence.
Noise reduction solutions have also emerged but face significant constraints. Many of the current models are trained on synthetic or limited datasets that do not reflect the range of clinically encountered noise types. Because synthetic datasets fail to capture the variability and complexity of actual catheterization lab imaging, AI models trained on such data may not generalize well to real world angiographic images, resulting in inconsistent or unreliable performance.

The prior art US20220175458A1 discloses AI-based image denoising systems for medical imaging. It acquires intra- and post-operative images during DBS procedures, but without coronary relevance or dynamic imaging scenarios. It develops a training dataset for DBS lead orientation, with synthetic noise and basic labels and lacks procedural, anatomical, and clinical depth for cardiac imaging. In contrast, the present invention provides a real-like dataset acquisition system that imitates coronary angiographic conditions using anatomically accurate phantoms, clinically relevant noise simulations, and actual Cath Lab protocols, enabling realistic ground-truth data acquisition that mirrors real patient procedures.

The prior art US20230162328A1 teaches machine learning-based methods relying on large-scale training data. It addresses noise suppression but focuses largely on computational frameworks and post-processing algorithms. While the document involves AI model training for image quality improvement, it relies on post-processing transformations and lacks any realistic coronary imaging setup or coronary acquisition method or physical noise simulation. In contrast, present solution creates imaging datasets in real time by replicating clinical conditions, including physical noise injection, real catheterization lab imaging modes, and structured metadata labelling thereby offering authenticity..

The prior art CN116531011A relates to X-ray imaging improvements using deep learning techniques and discusses reducing image noise in medical diagnostics. It suggests that datasets of noisy images may be useful for AI training but lacks detail on dataset generation protocols. The document focuses on software-based denoising using synthetically corrupted images and deep learning, without real-world coronary data acquisition or structured anatomical phantoms. It lacks specificity in imaging modality, procedure, and physical realism. In contrast, our invention provides a full-stack coronary imaging simulation—from phantom setup to imaging acquisition under various projection angles, SID distances, and magnification levels delivering a dataset that closely emulates actual angiographic procedures.

The prior art CA2391132A1 discusses general-purpose imaging phantoms to enhance diagnostic accuracy. It references multiple imaging setups but remains focused on clinical imaging datasets. While the document discusses general-purpose imaging phantoms, it lacks specificity for coronary anatomy, fluoroscopy-based data acquisition, and clinical imaging protocols. It is not intended for AI training or noise modeling. In contrast, our invention places the phantom on a PMMA slab to replicate body mass, varies SID to reflect patient geometry, and integrates labelling systems that capture all acquisition variables, making the dataset clinically actionable.

The prior art IN202311025908A describes AI-driven medical image enhancement workflows.It discloses the use of imaging phantoms for determining peripheral breast thickness without focus on coronary-specific anatomy. It hints at using training datasets for machine learning models but focuses on algorithmic techniques, not acquisition protocols. The document introduces a general concept of using phantom models for dataset generation but lacks specificity regarding cardiac anatomy or procedural fidelity for coronary angiography. It does not describe placement on a PMMA slab for tissue simulation. Our invention ensures anatomical realism using a custom coronary phantom and simulates clinical angiographic conditions in real time, producing datasets suitable for development, testing, and validation of imaging systems and diagnostic tools.

The prior art RU2555122C2 concerns X-ray image processing and diagnostic enhancement methods, including noise reduction and image clarity improvements. It generally addresses improving X-ray image quality for medical applications. The document introduces a general system for phantom-based simulation of imaging conditions but lacks the coronary-specific anatomical fidelity, real-world imaging modalities, and physical noise simulation techniques required for AI-based denoising. Our invention fills this critical gap by offering an end-to-end framework for generating coronary imaging datasets under real-like conditions using physical setups, imaging workflows, and structured image pairing.

The prior art JP2005253572A discloses imaging equipment and methods for obtaining medical images, including some X-ray imaging enhancements. It touches upon fluoroscopy and angiography but in the context of imaging devices and real patient data. The document presents a system for general X-ray image quality improvement using imaging adjustments, but does not involve a cardiac phantom, clinical coronary workflows, or physical noise injection. Our invention fills this critical gap by offering an end-to-end framework for generating coronary imaging datasets under real-like conditions using physical setups, imaging workflows, and structured image pairing.

Hence, traditional noise reduction techniques in coronary angiography, such as filtering or interpolation methods, often fail to preserve intricate vessel details and may introduce unwanted artifacts. Additionally, high-resolution imaging systems that could improve clarity come with high costs, increased storage needs, and additional radiation exposure.The present invention overcomes these shortcomings and providing a high-fidelity, real-like dataset acquisition system that supports the development of tools aimed at improving image quality, reducing diagnostic uncertainty, and enhancing patient outcomes.

OBJECTS OF THE INVENTION
The principal objective of the invention is to provide a method for generating angiographic imaging datasets that replicate real-world coronary angiography conditions, enabling improved noise reduction, image clarity, and diagnostic performance through machine learning.

Another objective of the invention is to provide an integrated system combining the phantom, noise simulation module, imaging system, and storage unit to streamline imaging dataset creation that replicate real-world coronary angiography conditions enabling improved noise reduction, image clarity, and diagnostic performance through machine learning.

Said and other objects of the present disclosure will be apparent to a person skilled in the art after consideration of the following summary of subject matter as claimed, detailed description taken into consideration with accompanying drawings in which preferred embodiments of the present disclosure are illustrated.

SUMMARY OF THE INVENTION
It is therefore a general aspect of the embodiments to provide a method for real-like imaging dataset generation imitating coronary angiographic conditions. The process begins with setting up a special heart model called an anthropomorphic phantom, which closely resembles the structure and appearance of a human heart, especially the coronary arteries. This phantom is built with artificial components such as vessels, stents, catheters, and wires that are externally placed on the phantom to mimic real-life procedures performed during coronary angiography and obtain clinical patient data. To make the images realistic, different types of noise (like blur, graininess, or scatter) are intentionally added to the imaging environment. This is done using physical techniques or materials that simulate the kind of image distortions that normally happen during actual heart imaging in hospitals. Using a real cardiac imaging system—like the ones found in catheterization labs—many images of the phantom are taken. These are captured from different angles, under different modes (such as low-dose or cine imaging), and at different zoom levels to cover a variety of imaging situations that doctors usually encounter. For every noisy image that’s captured, another image is taken under normal, clean conditions (for example, using frame averaging). The two images—noisy and clean—are paired together, so that both show the same scene but with and without noise. Each image pair is tagged with important information like what angle it was taken from, what imaging settings were used. This metadata ensures every image can be clearly understood and reused later. All these paired and labeled images are saved in a proper format—either in a database or a secure system. This method provides a systematic and clinically realistic way to generate high-quality angiographic image datasets that closely replicate what doctors see during real coronary procedures.

It is a further aspect of the embodiments clean reference images are created by using frame averaging. Frame averaging is a setting in X-ray machines that takes multiple images over time and combines them into one, reducing random noise while keeping the important anatomical features intact, giving machine learning models a trustworthy ground truth to compare against the noisy image.

It is another aspect of the embodiments to provide that each denoised image corresponds exactly to the noisy image is taken at the exact same angle, position, and configuration. This includes maintaining the same phantom orientation, imaging geometry, and magnification. This pixel-to-pixel alignment of the image pair means the AI model can precisely learn what parts of the image are noise and what parts are real, making the denoising process far more accurate and clinically dependable.

It is an aspect of the embodiments to provide that noise is introduced by adding salt or similar particulate solutes to the water surrounding the phantom. These particles slightly change how X-rays behave as they pass through, introducing realistic scatter and noise effects. This simulates unpredictable distortions seen in clinical settings, like patient movement or tissue variation, thereby enhancing the dataset’s realism and preparing the model for real-world challenges.

It is an aspect of the embodiments to provide that copper plates of different thicknesses are placed in the X-ray beam’s path (usually on the collimator). These metals cause attenuation and distortion similar to those created by actual medical devices inside the body. This step replicates visual disturbances caused by stents, catheters, and wires in real procedures, teaching the AI how to handle such complex image features without mistaking them for noise.

It is an aspect of the embodiments to provide that frame averaging is intentionally disabled during image capture. Without it, the resulting images contain more noise and grain—mimicking low-dose or emergency imaging scenarios. This helps the dataset reflect difficult imaging conditions, making the AI robust and effective even when operating under low-quality or time-sensitive imaging settings.

It is another aspect of the embodiments to provide that the images are captured at multiple projection angles — including Left Anterior Oblique (LAO), Right Anterior Oblique (RAO), cranial, caudal, and combinations thereby reflecting the full range of clinical viewing perspectives, and ensuring the dataset is representative of actual practice.

It is an aspect of the embodiments to provide that the images are captured using various modes like standard fluoroscopy, low-dose fluoroscopy, and cine imaging. Each mode has its own characteristics—fluoroscopy may be noisier, while cine offers clearer detail but different motion noise. The multiple imaging modes helps the model generalize across the full spectrum of machine settings it might encounter in real-world procedures.

It is an aspect of the embodiments to provide that multiple magnification levels (M0–M3) are used, from no magnification to high zoom. This affects image resolution, sharpness, and noise visibility—especially when looking at small or detailed vessel structures. The Training with multiple zoom levels ensures the AI system can denoise images whether the doctor is viewing a broad area or zooming in on fine details like vessel branches or stents.

It is an aspect of the embodiments to provide that the phantom simulation setup is placed in water and on a slab of PMMA (a plastic that mimics human body thickness). The source-to-detector distance (SID) is also varied to simulate different patient body sizes and setups. This simulates how X-rays interact with actual human tissue and body shapes, making the training data more realistic and the AI more applicable in daily clinical use.

It is another aspect of the embodiments to provide that each image pair is tagged with details like imaging angle, zoom level, SID. This helps track exactly how and when each image was created. The traceability and structured labeling enhance the training process, ensuring the AI model can relate noise patterns to imaging conditions crucial for safe clinical deployment.

It is an aspect of the embodiments to provide that, in addition to core imaging parameters, the metadata. Documenting the exact source and nature of noise enables AI models to associate specific distortion patterns with their physical causes, thereby enhancing the model’s ability to perform targeted and context-aware denoising with greater accuracy and reliability.

It is a further aspect of the embodiments to provide a system for Real-Like Imaging Dataset Generation Imitating Coronary Angiographic Conditions comprising the phantom which is a physical model of the human heart that includes artificial vessels, stents, guide wires, and catheters that are externally placed on the phantom to mimic a human heart and provide real-life clinical patient data. It closely mimics the structure and radiographic behavior of actual coronary anatomy, including how it appears under X-ray imaging. A noise simulation module responsible for introducing different types of image-degrading noise, such as scatter or geometric distortion. It can use mechanisms like copper plates, saline-based scatter media, or manipulation of imaging parameters to simulate real-life clinical image quality issues. An interventional X ray imaging system used in catheterization labs. It is configured to operate in multiple modes (e.g., fluoroscopy, cine, low-dose) and under different projection angles and zoom levels. It captures angiographic image sequences of the phantom under both clean and noisy imaging conditions. An image acquisition unit collects image data from the imaging system and organizes it into paired datasets. For every noisy image captured with noise added, it pairs it with a clean version taken under standard, optimal imaging conditions—without the added noise. This one-to-one pairing forms the basis for supervised machine learning. A processing module adds important metadata to each image pair. The metadata includes details like the projection angle, magnification, imaging mode. This structured labelling ensures that every image is traceable and usable for high-quality model training. Lastly, a data storage unit where the labeled paired datasets are saved. The storage format ensures that the image pairs and their metadata are easily retrievable and compatible with AI training frameworks. The result is a highly realistic, clinically relevant, and fully traceable dataset creation platform. This system enables the development of AI models that can effectively reduce noise in coronary angiographic images without compromising diagnostic accuracy.

BRIEF DESCRIPTION OF DRAWINGS
The detailed description is described with reference to the accompanying figures. In the figures, the left-most digit(s) of a reference number identify the figure in which the reference number first appears. The same numbers are used throughout the drawings to reference like features and modules.

Figure 1 describes a flow chart of method steps for generating angiographic imaging datasets that replicate real-world coronary angiography conditions.
Figure 2 depicts a block diagram to illustrating the internal system architecture for carrying out the method of Figure 1.
Figure 3 illustrates examples of angiographic images that provide a direct comparative analysis between noisy and denoised images captured under different imaging conditions.

The figures depict embodiments of the present subject matter for the purposes of illustration only. A person skilled in the art will easily recognize from the following description that alternative embodiments of the method and composition illustrated herein may be employed without departing from the principles of the disclosure described herein.

DETAILED DESCRIPTION
The best and other modes for carrying out the present invention are presented in terms of the embodiments, herein depicted in drawings provided. The embodiments are described herein for illustrative purposes and are subject to many variations. It is understood that various omissions and substitutions of equivalents are contemplated as circumstances may suggest or render expedient, but are intended to cover the application or implementation without departing from the spirit or scope of the present invention. Further, it is to be understood that the phraseology and terminology employed herein are for the purpose of the description and should not be regarded as limiting. Any heading utilized within this description is for convenience only and has no legal or limiting effect.

The terms “a” and “an” herein do not denote a limitation of quantity, but rather denote the presence of at least one of the referenced items.

The terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process or method that comprises a list of steps does not include only those steps but may include other steps not expressly listed or inherent to such a process or method. Similarly, one or more sub-systems or elements or structures or components preceded by "comprises... a" does not, without more constraints, preclude the existence of other, sub-systems, elements, structures, components, additional sub-systems, additional elements, additional structures or additional components. Appearances of the phrase "in an embodiment", "in another embodiment" and similar language throughout this specification may, but not necessarily do, all refer to the same embodiment.

For the purpose of promoting an understanding of the principles of the invention, reference will now be made to the embodiment illustrated in the figures and specific language will be used to describe them. It will nevertheless be understood that no limitation of the scope of the invention is thereby intended. Such alterations and further modifications in the illustrated system, and such further applications of the principles of the invention as would normally occur to those skilled in the art are to be construed as being within the scope of the present invention.

It will be understood by those skilled in the art that the foregoing general description and the following detailed description are exemplary and explanatory of the invention and are not intended to be restrictive thereof.

Unless otherwise defined, all technical and scientific terms used herein have the same meanings as commonly understood by those skilled in the art to which this invention belongs.

The method and system provided herein are only illustrative and not intended to be limiting. Embodiments of the present invention will be described below in detail with reference to the accompanying figures.
The present invention provides a novel approach for real-like imaging dataset generation that imitates coronary angiographic conditions in a controlled environment. In a controlled catheterization lab setting, an anthropomorphic cardiac phantom replicating realistic coronary vessel anatomy is imaged under conditions that closely mirror those encountered in clinical practice. Multiple projection angles (including left anterior oblique, right anterior oblique, cranial, and caudal views) and imaging modes (such as fluoroscopy, low-dose fluoroscopy, and cine) are used to capture a diverse set of angiographic images, ensuring that the dataset reflects the range of perspectives and exposure settings typical of real-world coronary imaging.
To create robust and representative training data, noise is deliberately introduced through controlled, noise injection mechanisms. These techniques generate a comprehensive spectrum of temporal and geometric noise artifacts, effectively simulating the visual challenges faced in clinical angiography, such as scatter, motion blur, and attenuation distortions. For each imaging configuration, both a noisy image and a corresponding denoised counterpart—captured under standard imaging conditions with frame averaging enabled—are recorded, forming a paired dataset that presents the same anatomical scene in noisy and clean states.
Each image pair is then meticulously labeled with metadata detailing key acquisition attributes, including projection angle, imaging mode, magnification level. This structured and traceable labeling ensures the dataset is as close to the actual conditions encountered in clinical practice as possible and every image in the dataset can be tracked, classified, and contextualized for model training. By providing a dataset that authentically reflects the complexity and variability of coronary angiography without resorting to synthetic data or costly hardware upgrades, this invention creates a powerful foundation for training machine learning models. The result is a cost-effective, clinically relevant pathway to enhancing angiographic image quality—enabling AI systems to reduce noise while preserving fine vascular details, ultimately improving diagnostic accuracy and patient outcomes.

Fig. 1 illustrates a flow chart of the method (100) for imaging dataset generation that imitates coronary angiographic conditions in a controlled environment. The process begins with simulating (101) a realistic coronary imaging scenario through the use of an anthropomorphic cardiac phantom which is a physical model designed to replicate the shape, structure, and imaging behavior of the human heart and its coronary vessels. This phantom is specifically engineered to replicate the structural characteristics i.e physical form and arrangement of the coronary anatomy, including Coronary arteries (left main, LAD, circumflex, RCA), Branching patterns, Vessel diameters and curvatures, as well as the radiographic characteristics appears under X-ray imaging, including X-ray attenuation properties, visibility of metallic components, contrast medium behavior during angiography. This phantom is equipped with key anatomical components such as a vessel, stent, catheter, guide wire. The phantom comprises any combination of these components (or even just one) to simulate vascular interventions. These embedded structures replicate real human arteries, their vascular interventions and add complexity to the images, allowing the captured angiograms to closely mirror those seen in clinical coronary angiography. By faithfully modeling these anatomical features, the phantom ensures that the images reflect true clinical challenges, including the visualization of stents and wires within narrow or branching vessels.
In one embodiment, the invention further refines the simulation of the coronary imaging scenario by recreating the physical conditions of a clinical catheterization laboratory through specific phantom setup techniques. The anthropomorphic cardiac phantom is fully immersed in water, a well established approach in radiology for replicating the attenuation and scattering behavior of human soft tissue under X ray imaging. Water immersion is a well-established imaging technique to simulate the attenuation properties of human soft tissue and its effect on X-ray penetration. This step creates a uniform imaging medium, eliminates extraneous environmental interference, and standardizes image quality. By replicating how X-rays interact with real tissue and blood vessels, the water immersion allowing the dataset to accurately reflect how coronary vessels appear in clinical angiography by eliminating unwanted external influences.
Additionally, the phantom was placed on a 20 mm Polymethyl Methacrylate (PMMA) slab, a material widely used in medical imaging to simulate the density and thickness of a patient’s body. This PMMA layer ensures that the images capture the same X ray absorption patterns and anatomical depth challenges encountered during real coronary angiography procedures.
Additionally, the embodiment introduces variations in Source to Detector Distance (SID), ranging from 995 mm to 1082 mm during image acquisition. By deliberately changing the SID, the imaging system captures differences in magnification, sharpness, and background noise, mirroring the variability caused by different patient sizes, table heights, or equipment positioning in actual clinical settings. These variations are key to generating a diverse dataset that exposes full spectrum of imaging conditions they will encounter in practice. Hence, water immersion, PMMA slab placement, and SID variation creates a high fidelity simulation environment reflect the nuances and variability of true coronary angiography and train model on this data to perform reliably across a wide range of clinical scenarios.
In the second step of the method (100), controlled noise (102) is deliberately introduced into the imaging environment to ensure that the angiographic dataset authentically represents the kinds of image degradations encountered in a clinical catheterization lab. Rather than relying on naturally occurring imperfections alone, the invention takes a proactive approach, using physical noise-inducing mechanisms to replicate clinically relevant artifacts. This deliberate “noise injection” provides a spectrum of distortions that mimic the challenges faced by radiologists and cardiologists during real angiographic imaging — including scatter, attenuation, and temporal and geometric variations. By creating noise in a controlled and measurable way, the resulting dataset becomes an invaluable foundation for training AI models to recognize and reduce these disruptions while preserving critical vascular details. In the second step of the method (100), controlled noise is deliberately introduced into the imaging environment to ensure that the angiographic dataset authentically represents the kinds of image degradations encountered in a clinical catheterization lab. Rather than relying on naturally occurring imperfections alone, the invention takes a proactive approach, using physical noise-inducing mechanisms to replicate clinically relevant artifacts. In this mechanism, there is delebrate introduction of physical materials into the imaging environment to create scatter, attenuation, and other distortions. This deliberate “noise injection” provides a spectrum of distortions that mimic the challenges faced by radiologists and cardiologists during real angiographic imaging including scatter, attenuation, and temporal and geometric variations. By creating noise in a controlled and measurable way, the resulting dataset becomes an invaluable foundation for training models to recognize and reduce these disruptions while preserving critical vascular details.
In one embodiment, a pinch of salt or similar particulate solutes is added to the immersion water surrounding the anthropomorphic phantom. This seemingly simple step creates irregularities in the imaging medium, altering how X rays pass through and scatter. The resulting subtle distortions replicate the natural noise variations caused by real-world factors such as minor patient movements, tissue inconsistencies, or equipment fluctuations. This “salt-based” noise creates fine-grain scatter artifacts, which are difficult for conventional denoising algorithms to handle but vital for building models that can learn to distinguish between true anatomical detail and random background distortions.
In another embodiment, copper plates of 1 mm and 5 mm thickness are strategically placed on the collimator of the Cath lab imaging system. These copper inserts are used to deliberately block and distort portions of the X ray beam, introducing geometric and temporal noise similar to the interference caused by medical devices such as stents, catheters, or guide wires. The copper plates create attenuation artifacts and misalignment distortions, replicating the unpredictable challenges clinicians face in reading angiograms. By varying the plate thickness and placement, a range of noise intensities and patterns is generated, enriching the dataset with both mild and severe noise scenarios.
Further, to create a clinically realistic and challenging angiographic dataset, the invention employs a broad range of noise inducing mechanisms, including imaging parameter manipulations, environmental or setup alterations, and optional digital augmentation, all designed to replicate the diverse ways image quality can degrade in a real catheterization laboratory.
In terms of imaging parameter manipulations, noise is deliberately generated by adjusting how the imaging system operates — for example, disabling frame averaging to remove inherent smoothing, modifying pulse width to increase quantum noise, reducing radiation dose, or slightly misaligning the X ray source and detector to create geometric distortions.
In the category of environmental or setup alterations, noise is introduced by changing the physical conditions surrounding the phantom, such as creating air gaps to induce scatter, placing non uniform background materials (e.g., foam or plastic sheets) to produce uneven attenuation, or simulating motion to mimic patient breathing or heartbeat.
The dataset may also optionally include digital augmentation techniques, where controlled amounts of synthetic noise — such as Gaussian, Poisson, or Rician distributions — are added post acquisition to supplement the range of noise profiles captured physically.
By incorporating this broad spectrum of noise inducing approaches, the invention ensures that the resulting dataset spans a wide spectrum of clinically relevant noise conditions in clinical angiography, resulting in robust, generalizable noise reduction models that preserve fine vascular detail and support more accurate diagnosis and treatment planning.
Once the phantom environment has been prepared and noise has been introduced, the next step of the method (100) involves capturing a multiple noisy angiographic images (103) using a state of the art catheterization lab (Cath lab) imaging system. This acquisition step is central to the invention, as it ensures that the dataset authentically reflects how coronary angiography is performed in clinical settings. The imaging system is operated across a range of realistic configurations such as projection angles, imaging modes, and magnification levels etc. to mirror the conditions cardiologists and radiologists encounter when diagnosing coronary artery disease. Each imaging session collects noisy angiographic image sequences directly from the phantom under controlled but clinically relevant circumstances.
In one embodiment, images are captured using multiple projection angles to replicate the standard practice of visualizing coronary vessels from different viewpoints. These angles include Left Anterior Oblique (LAO), Right Anterior Oblique (RAO), cranial, and caudal orientations, as well as combinations thereof. Capturing the phantom from these varying angles produces images that reflect the full spectrum of anatomical perspectives used by cardiologists in real procedures. This ensures that the dataset is not limited to a single viewpoint but instead represents the true multi angle imaging workflow of clinical coronary angiography, which is essential for developing models that can denoise images regardless of the viewing angle.
In another embodiment, the invention employs multiple imaging modes during acquisition to further expand the dataset’s realism. Modes include fluoroscopy, low dose fluoroscopy, and cine imaging, each introducing distinct noise patterns and exposure characteristics. For example, low dose fluoroscopy inherently produces grainier, noisier images, while cine imaging generates higher resolution sequences but with different temporal noise profiles. Including all these imaging modes ensures the dataset captures a comprehensive range of clinical scenarios. Each of these imaging modes introduces distinct noise characteristics, making them valuable for creating a comprehensive dataset that reflects real-world variability. This variability is critical to developing machine learning models that can reduce noise across a wide range of conditions while preserving fine vascular details essential for accurate diagnosis.

In a further embodiment, the data acquisition step includes the use of multiple magnification levels, ranging from M0 (no magnification) to M1, M2, and M3 (progressively higher zoom levels). These magnification levels simulate how cardiologists zoom in during procedures to better visualize intricate or small vessels. Each magnification level alters image characteristics, including resolution, apparent vessel size, and the visibility of noise. Hence, each magnification level introduces specific challenges, particularly regarding noise characteristics, which are crucial for training the machine learning models to handle varying magnification levels while preserving image clarity.
In addition to projection angles, imaging modes, and magnification levels, other imaging configurations can be varied to capture the full spectrum of noise conditions encountered in coronary angiography. These include frame rate adjustments (e.g., 7.5, 15, or 30 frames per second) to introduce different temporal resolutions and motion artifacts, exposure parameters such as kilovoltage peak (kVp) and tube current (mA) to simulate dose dependent noise, and pulse width modifications that influence quantum noise levels. Further variations can be achieved by altering detector settings (such as pixel binning or resolution mode), changing collimation to affect scatter and edge artifacts, and adjusting table height or gantry tilt to replicate geometric distortions seen in practice. Additional configurations can include manipulating contrast injection variables, using or removing anti scatter grids, or switching detector modes between high resolution and high sensitivity settings. Collectively, these adjustments ensure the dataset captures a diverse range of clinically relevant imaging conditions, enabling models to generalize effectively across all real world scenarios.
In the fourth step of the method (100), the invention focuses on generating paired image datasets (104) that will form the foundation for training AI-driven noise reduction models. After noisy and denoised images of the anthropomorphic cardiac phantom have been captured under various imaging conditions, the method (100) systematically pairs each noisy angiographic image with its corresponding denoised counterpart. Each pair represents the exact same anatomical scene, ensuring that any visual difference between the two images is solely due to the presence or absence of noise. This pairing process is critical because it provides the “before and after” reference needed for supervised machine learning—allowing machine learning models to learn how to remove noise while preserving subtle vascular structures. The creation of these pairs ensures that the dataset is structured, traceable, and meaningful, rather than a loose collection of unrelated images.
In one embodiment, the denoised reference images are acquired by enabling frame averaging on the imaging system during acquisition under standard clinical settings. Frame averaging is a technique that combines multiple exposures of the same scene to produce an image with reduced temporal noise and enhanced clarity. By using frame averaging, the invention generates high quality reference images that serve as reliable “ground truth” examples for the models. These clean images help distinguishing real anatomical detail from noise distortions, ensuring that noise reduction does not inadvertently remove delicate vascular features.
In another embodiment, every denoised image corresponds precisely to a noisy image captured under the same anatomical positioning and geometric configuration. This means that the phantom orientation, imaging angle, source to detector distance, magnification, and other geometric parameters remain identical between the noisy and denoised acquisitions—except for the applied noise-inducing conditions. This one-to-one correspondence guarantees pixel level alignment between noisy and denoised versions of the same image, which is essential for creating paired datasets that an AI model can use to learn precise noise removal mappings. By relying on these paired noisy/denoised images rather than costly hardware upgrades or increased radiation doses, this method (100) offers a cost-effective, safe, and clinically relevant route to enhanced angiographic image quality.
In the fifth step of the method (100), once the noisy and denoised image pairs are generated, the invention focuses on meticulous labelling (105) of each dataset to ensure the information is structured, traceable, and directly usable for training of machine learning models. Each paired image set is annotated with key acquisition details, creating a rich layer of metadata that links every image. This structured labelling transforms a raw collection of images into a highly organized training resource.
In one embodiment, the metadata labelling includes a comprehensive set of acquisition attributes, ensuring that every relevant imaging parameter is documented. Each image pair is tagged with its projection angle (e.g., LAO, RAO, cranial, caudal), imaging mode (fluoroscopy, low-dose fluoroscopy, or cine), and magnification level (M0–M3). Additional metadata fields identify the source-to-detector distance (SID) used during capture, and the timestamp of acquisition.
This rich metadata ensures that every image can be tracked, classified, and recalled with full context, enabling the models to differentiate noise profiles and understand the imaging conditions that produced them. As a result, the models are trained not just to “remove noise,” but to intelligently suppress different types of artifacts while retaining fine anatomical details just as a cardiologist would expect when interpreting real clinical images.

The final step of the method (100) involves storing (106) the labelled paired datasets in a secure and organized manner to ensure they are readily usable for training machine learning models configured to reduce or suppress noise in coronary angiographic images. After each noisy and denoised image pair has been meticulously labelled with metadata describing its imaging conditions, the complete dataset is compiled into a structured, digital archive. This storage step is not simply a data dump, it involves arranging the image pairs and their metadata into a computer readable format (such as a database or cloud based repository) that allows for easy retrieval, sorting, and indexing.

By storing the dataset in this structured format, the invention ensures that the paired images and their associated labels are traceable and interoperable with training pipelines. Machine learning models can access the data efficiently for supervised learning, using the noisy image as input and the denoised counterpart as the target output. This organized storage also enables future expansion of the dataset new image pairs from additional noise conditions, imaging modes, or phantom configurations can be seamlessly added without disrupting the dataset’s structure.

Ultimately, this storage step transforms the labelled paired images into a living resource for development. By maintaining the dataset in an accessible and standardized form, it supports the training, testing, and validation of machine learning models that learn to remove noise from coronary angiographic images, enhancing image clarity and diagnostic accuracy without the need for costly hardware upgrades or increased patient radiation exposure. These datasets are structured to reflect clinical realism and are intended for use in developing and validating medical image enhancement solutions, including—but not limited to—AI-based noise reduction techniques.

Hence, the proposed solution provides a ground breaking and clinically meaningful approach to the acquisition, preparation, and labelling of a specialized dataset designed for noise reduction in coronary angiography. This invention makes it possible to produce clearer angiographic images without relying on costly high end imaging systems or subjecting patients to increased radiation doses. The approach represents a significant advancement in medical imaging technology, ultimately improving diagnostic accuracy, aiding interventional planning, and contributing to better patient outcomes and more efficient healthcare delivery.

Figure 2 depicts a block diagram depicting the internal system architecture (200) for real-like imaging dataset generation imitating coronary angiographic conditions. Hardware and software components into a coordinated platform, designed to emulate the coronary imaging environment and produce high-fidelity, paired noisy and denoised angiographic image datasets. These datasets are structured to reflect clinical realism and are intended for use in developing and validating medical image enhancement solutions, including—but not limited to—AI-based noise reduction techniques.
At the core of the system (200) is an anthropomorphic cardiac phantom (201), which is configured to replicate the structural and radiographic characteristics of human coronary anatomy. The phantom includes artificial components such as vessels, stents, guide wires, and catheters, placed externally on the phantom to simulate real-world vascular interventions and imaging challenges and mimic real-life clinical patient data. This phantom (201) provides a clinically accurate subject for image acquisition, ensuring that the resulting dataset mirrors the complexity of coronary angiography.
A noise simulation module (202) is connected to the phantom environment and is configured to introduce one or more noise-inducing mechanisms (202a) into the imaging setup. This controlled noise injection (202a) enables the system to generate images with a broad spectrum of temporal and geometric artifacts, representative of those seen in actual catheterization labs.
An imaging system (203), such as an interventional cardiac X ray unit operable in a catheterization laboratory, is configured to capture angiographic image sequences of the phantom. The system (203) operates under a variety of imaging configurations, including multiple projection angles (203a), imaging modes (203b) (fluoroscopy, low dose fluoroscopy, cine), and magnification levels (203c). These variations ensure that the acquired images reflect the full range of clinical conditions and viewpoints used by cardiologists.
The paired image datasets (noisy and denoised) is captured within the acquisition unit (204). For each imaging configuration, the acquisition unit (204) produces a noisy image (204a) (captured under noise inducing conditions) and a corresponding denoised image (204b) (captured under standard imaging conditions). These paired images (204a, 204b) are perfectly aligned, depicting the same anatomical scene, and provide the “before and after” structure necessary for supervised training.
The processing module (205) then takes over, labelling each paired dataset with comprehensive metadata. This metadata (205a) includes details such as projection angle, imaging mode, magnification level, source to detector distance. By embedding this information, the processing module (205) ensures that every image pair is traceable and searchable, making the dataset highly structured and directly usable for machine learning model development.
Finally, a data storage unit (206) serves as the repository for the labelled paired datasets. The storage unit maintains the data in a secure, organized, and computer readable format, enabling it to be easily accessed, expanded, and integrated into machine learning workflows. The stored datasets are then used to train the models to reduce or suppress noise in coronary angiographic images without erasing subtle vascular details.
Fig. 3 illustrates examples of angiographic images that provide a direct comparative analysis between noisy and denoised images captured under different imaging conditions imitating coronary angiographic conditions.
On the left side of the figure, representative noisy images are shown. These were acquired using an anthropomorphic cardiac phantom under controlled noise inducing conditions. Techniques such as salt infused water immersion and the placement of copper plates on the collimator were applied, and in some cases, frame averaging was disabled. Together, these methods generated a wide spectrum of temporal and geometric noise artifacts, including scatter, graininess, and attenuation distortions. These noisy images accurately mimic the quality degradations commonly encountered in a real clinical catheterization laboratory when lower dose imaging settings, patient motion, or physical obstructions affect image clarity.
On the right side of the figure, corresponding denoised images are shown. These were captured under standard imaging conditions with frame averaging enabled and without the additional noise inducing mechanisms. These clean reference images preserve fine vascular detail and provide a “ground truth” baseline against which the noisy images can be compared.
By presenting noisy and denoised images side by side, Figure 2 demonstrates how each image pair represents the same anatomical scene of the phantom but under two different imaging scenarios — one degraded by controlled noise and one captured in optimal conditions. This contrast forms the foundation of the paired dataset: the noisy image serves as the input for the machine learning model, while the denoised counterpart serves as the target output for supervised training.
These image comparisons also highlight the breadth of imaging variability captured in the dataset. Images were obtained across multiple projection angles, imaging modes, and magnification levels. This variety ensures that the model learns to handle noise from diverse clinical situations while preserving subtle anatomical details, such as small vessel branches or stent outlines.
In summary, Fig. 3 creating structured pairs of noisy and denoised angiographic images that reflect real world imaging variability, providing the essential “before and after” examples for training the machine learning models to reduce noise without erasing crucial diagnostic information.
, Claims:I/We Claim:
1. A method (100) for real-like imaging dataset generation imitating coronary angiographic conditions, the method (100) comprising:
-simulating (101) a coronary imaging scenario using an anthropomorphic phantom which is designed to replicate the structural and radiographic characteristics of human coronary anatomy, the phantom comprises artificial vessels, stents, guide wires, and catheters, configured to simulate vascular interventions;
-introducing noise (102) including temporal noise and geometric noise, faced in real catheterization lab imaging into the imaging environment by incorporating one or more noise-inducing mechanisms to obtain simulated noise that mimics image degrading effects and conditions occur during coronary angiography in a clinical environment;
-capturing (103) a plurality of noisy angiographic image sequences of the cardiac phantom, using interventional cardiac imaging system operable in catheterization laboratories, the image acquisition being performed under a range of imaging configurations includes one or more of projection angles, imaging modes, and magnification levels;
-generating (104) paired image datasets for each imaging configuration, each dataset comprising:
a noisy image of the cardiac phantom, acquired under noise-induced mechanisms; and
a corresponding denoised image of the cardiac phantom, acquired under standard imaging conditions;
-labelling (105) each paired image dataset with metadata to identify the imaging configuration ; and
-storing (106) the labelled paired imaging datasets that emulate coronary angiographic conditions for reduction or suppression of noise in coronary angiographic images.

2. The method (100) as claimed in claim 1, wherein the denoised images are acquired by enabling frame averaging at standard imaging settings.

3. The method (100) as claimed in claim 2, wherein each denoised image is corresponds to a noisy image acquired under the same anatomical positioning and geometric configuration.

4. The method (100) as claimed in claim 1, wherein the noise-inducing mechanisms comprises adding a predefined amount of salt or other particulate solutes to the immersion water to simulate induced noise variations.

5. The method (100) as claimed in claim 1, wherein the noise-inducing mechanisms comprises placing copper plates of predefined thickness in the X-ray beam path such as on the collimator, to introduce noise variations.

6. The method (100) as claimed in claim 1, wherein the noise-inducing mechanisms comprise disabling frame averaging during image acquisition to generate noisy image sequences.

7. The method (100) as claimed in claim 1, wherein the projection angles comprises atleast one of Left Anterior Oblique (LAO), Right Anterior Oblique (RAO), cranial, caudal and the combination thereof to simulate various clinical views of coronary vasculature.

8. The method (100) as claimed in claim 1, wherein the imaging modes include at least one of fluoroscopy, low-dose fluoroscopy, and cine imaging each mode, contributing distinct noise characteristics.

9. The method (100) as claimed in claim 1, wherein the magnification levels comprise a plurality of zoom settings selected from M0 (no magnification), M1, M2 and M3 (high magnification), wherein each level simulates different clinical magnification condition .

10. The method (100) as claimed in claim 1, wherein simulating (101) the coronary imaging scenario further comprises immersing the anthropomorphic phantom in water to replicate the attenuation properties of human soft tissue during X-ray imaging; positioning the phantom on a slab of polymethyl methacrylate (PMMA) having a predefined thickness to simulate patient body mass and anatomical depth; and varying the source-to-detector distance (SID) during image acquisition to emulate real-world clinical imaging variability across patient sizes and table setups.

11. The method (100) as claimed in claim 1, wherein the metadata indicative of one or more imaging parameters, including projection angle, imaging mode, magnification level, source-to-detector distance (SID), and acquisition timestamp, thereby enabling traceability of image.

12. The method (100) as claimed in claim 1, wherein labelling (105) each paired image dataset with metadata comprises annotating each pair with projection angle, imaging mode, magnification level, source-to-detector distance (SID), and acquisition timestamp.

13. A system (200) for real-like imaging dataset generation imitating coronary angiographic conditions, the system (200) comprising:
• an anthropomorphic cardiac phantom (201), configured to simulate the structural and radiographic characteristics of human coronary anatomy, the phantom comprising artificial vessels, stents, guide wires, and catheters for replicating vascular interventions;
• a noise simulation module (202), configured to introduce one or more noise-inducing mechanisms (202a) into the imaging environment;
• an imaging system (203) operable in a catheterization laboratory, configured to acquire angiographic image sequences of the phantom under a plurality of imaging configurations, the imaging configurations comprising one or more of projection angles (203a), imaging modes (203b), and magnification levels (203c);
• an acquisition unit (204), configured to generate, for each imaging configuration, a paired image dataset comprising a noisy image (204a), acquired under noise-induced mechanisms and a corresponding denoised image (204b), acquired under standard imaging conditions to capture noise characteristics;
• a processing module (205), configured to label each paired image dataset with metadata (205a) indicative of the their imaging configuration; and
• a data storage unit (206), configured to store the labelled paired image datasets that emulate coronary angiographic conditions for reduction or suppression of noise in coronary angiographic images.