Method for nuclear image enhancement using long-lived isotope-labeled radiotracers and deep learning

Abstract

The current invention relates to a method of nuclear imaging, wherein said method uses an isotope-labeled radiotracer, wherein said isotope has a half-life of at least (750) minutes, said method comprises acquiring a first image by means of a nuclear imaging method of a target area, and obtaining a second image from said first image, wherein said second image is obtained by improving said first image quality by means of a deep learning model.

Description

[0001] METHOD FOR NUCLEAR IMAGE ENHANCEMENT USING LONG-LIVED

[0002] ISOTOPE-LABELED RADIOTRACERS AND DEEP LEARNING

[0003] FIELD OF THE INVENTION

[0004] The field of the invention pertains to methods of nuclear imaging, specifically involving the use of isotope-labeled radiotracers with a half-life of at least 750 minutes. The current invention also pertains to computer-implemented method for improving nuclear images obtained via nuclear imaging methods such as a PET scan.

[0005] BACKGROUND

[0006] Nuclear imaging is a pivotal tool in medical diagnostics, enabling the visualization of physiological processes within the body through the use of isotope-labeled radiotracers. These radiotracers emit radiation detectable by imaging devices such as positron emission tomography (PET) or single photon emission computed tomography (SPECT) scanners. However, current methodologies face significant challenges that compromise the quality and reliability of the images produced. Novel PET radiotracers, although promising, encounter obstacles in broader adoption due to high costs and inconsistent image quality.

[0007] One major challenge is the limited half-life of certain isotopes, which constrains the duration and flexibility of imaging sessions. Additionally, PET images using long- lived radiotracers often suffer from low quality, marked by high noise levels and poor resolution, hindering accurate diagnosis and necessitating further processing to enhance image clarity. Traditional image enhancement techniques frequently fall short of addressing these deficiencies, thus creating a demand for more advanced methods.

[0008] Deep learning-based denoising presents a potential solution, yet it typically requires large training datasets that are often unavailable for new tracers. The selection of appropriate radiotracers and the optimization of imaging protocols also remain ongoing challenges. Therefore, there is an urgent need for innovative solutions that can improve the quality of nuclear imaging while accommodating the inherent limitations of current technologies.

[0009] SUMMARY OF THE INVENTION The present invention relates to a method of nuclear imaging that leverages isotope-labelled radiotracers according to claim 1. The method involves acquiring a first image of a target area using a nuclear imaging technique and subsequently obtaining a second image by enhancing the first image's quality through a deep learning model. This deep learning model can be a convolutional neural network (CNN) or an ensemble 3D U-Net architecture, which includes at least two models trained on datasets with varying noise levels. The model may sample 3D volumes of voxels from high-signal regions to improve image quality. In an embodiment, the isotopes used can be selected from 89Zr, 64CU, 1241, 52Mn, or 55Co. In an embodiment the first image is typically of low-grade quality. The nuclear imaging scan can be a PET scan, and the radiotracer can be an isotope-labelled biological agent such as a cell (e.g., a T cell), peptide, protein, or antibody. Specifically, the biological agent can be an antibody, preferably a nanobody, which may be able to bind to specific targets such as durvalumab or other specified markers. The invention offers several advantages, including earlier detection of abnormalities, improved spatial resolution, more precise targeting of treatments, and enhanced diagnostic accuracy. Additionally, the method allows for lower doses of radiotracers, reducing patient exposure to radiation, and provides robust imaging under various conditions due to the ensemble 3D U-Net architecture.

[0010] In a second aspect, the current invention also discloses a method according to claim 12, which is a computer-implemented method for analyzing and / or improving a nuclear image. This method involves receiving a first nuclear image, acquired using an isotope-labeled radiotracer with a half-life of at least 750 minutes, and obtaining a second image by enhancing the quality of the first image through a deep learning model. The deep learning model employed can be an ensemble 3D U-Net architecture comprising at least two models, each trained on datasets characterized by different levels of noise. The output of the models may be adaptively weighted based on the input image characteristics or the output of computer vision filters applied to the input image or model outputs. This dynamic weighting process ensures optimal image enhancement tailored to the specific imaging conditions. The invention also includes the ability to utilize the deep learning model to improve images of radiotracers that were not included in the training dataset, thereby demonstrating robustness and generalizability across different nuclear imaging scenarios. The method offers significant benefits in terms of diagnostic accuracy, reduced radiation exposure, and the ability to enhance image quality for both known and novel radiotracers, ensuring it remains applicable to a wide range of nuclear imaging applications. In a third aspect, the invention includes a method for denoising positron emission tomography (PET) images according to claim 19, which involves training a deep learning model specifically for enhancing PET images. The method comprises the steps of: (a) training the deep learning model on PET images that exhibit various levels of noise, including those reconstructed from list-mode raw data at reduced radioactive event counts, and (b) applying the trained model to PET images obtained using novel radiotracers that were previously unseen by the model, thereby enhancing image quality.

[0011] DESCRIPTION OF FIGURES

[0012] Figure 1 shows a schematic overview of a possible embodiment of a method according to the current invention. The embodiment shows a method for enhancing the low-grade quality input images with deep learning employing an ensemble 3D U-Net architecture. Furthermore, the embodiment visualizes the dynamic output weighting of the separate U-Net model outputs, such that the result is a series of enhanced images.

[0013] Figure 2 shows a schematic overview of a possible embodiment of a method according to the current invention. In this embodiment a low-quality image, obtained with a long-lived tracer is enhanced by a deep learning model, trained on short-lived radiotracer data, such that an enhanced version of the image is obtained.

[0014] Figure 3, 4 are described in the example sections.

[0015] DETAILED DESCRIPTION OF THE INVENTION

[0016] The current invention relates to methods of nuclear imaging employing deep learning models.

[0017] Definitions

[0018] The term "nuclear imaging" refers in the present invention to a medical imaging technique that uses radioactive substances to visualize and diagnose various conditions within the body. This technique includes but is not limited to positron emission tomography (PET) and single-photon emission computed tomography (SPECT).The term "isotope-labelled radiotracer" as used herein to refer to a compound that includes a radioactive isotope, which is used to trace the presence and distribution of the compound within a biological system. The isotope is chosen based on its half-life and compatibility with the imaging technique used. The term "half-life" refers to the time required for half of the radioactive atoms in a sample to decay. In the context of this invention, the isotope has a half-life of at least 750 minutes. The measurement of half-life should preferably follow ISO 11929:2019.

[0019] The term "first image" refers to the initial image acquired by a nuclear imaging method of a target area. This image may be of low-grade quality due to various factors such as noise and resolution.

[0020] By "second image" is meant the image obtained from the first image, wherein the quality of the first image is improved through the application of a deep learning model.

[0021] By "low-grade quality image" is meant an image with a signal-to-noise ratio below 3, where the signal-to-noise ratio is calculated by dividing the mean standard uptake value (SUVmean) by the standard deviation in standard uptake value, calculated in a uniform, spheric region of interest (ROI) in the liver with a diameter of at least 3 centimeters.

[0022] The term "deep learning model" refers to a type of machine learning model that uses neural networks with multiple layers to analyze and learn from data. In this invention, the deep learning model is used to enhance the quality of the first image. The term "convolutional neural network (CNN)" refers to a class of deep learning models particularly effective for image processing tasks. It uses convolutional layers to automatically and adaptively learn spatial hierarchies of features from input images.

[0023] The term "ensemble 3D U-Net architecture" refers to a specific type of deep learning model that consists of at least two 3D U-Net models. These models are trained on datasets characterized by differing levels of noise to improve image quality. Such ensemble 3D U-Net architecture is schematized in Figure 1.

[0024] The term "3D volumes of voxels" refers to three-dimensional units of graphic information that represent a value on a grid in three-dimensional space. In this invention, the model samples these 3D volumes from high-signal regions to enhance image quality. The term "PET scan" refers to positron emission tomography, a nuclear imaging technique that produces three-dimensional images of functional processes in the body.

[0025] The term "radiotracer" refers to a radioactive compound used in nuclear imaging to visualize and diagnose conditions within the body. Examples include isotope-labeled biological agents such as cells (e.g., T cells), peptides, proteins, or antibodies.

[0026] The term "antibody" refers to a protein produced by the immune system that can specifically bind to a target antigen. In this invention, the antibody may be a nanobody, which is a smaller, single-domain antibody.

[0027] The terms "durvalumab", "crefmirlimab berdoxam", refer to a specific antibody used in cancer immunotherapy. Throughout this text, these terms are used as nonlimiting examples of monoclonal antibodies.

[0028] Detailed description

[0029] In a first aspect, the present invention relates to a method of nuclear imaging that utilizes an isotope-labelled radiotracer. The isotope employed in this method has a half-life of at least 750 minutes. This method comprises acquiring a first image of a target area through a nuclear imaging technique and subsequently obtaining a second image from the first image. The second image is generated by improving the quality of the first image using a deep learning model.

[0030] Prior to acquiring said first image of a target area, the administration of the isotopelabelled radiotracer to the subject can be performed through various routes, such as intravenous injection, oral ingestion, or inhalation, depending on the specific radiotracer and the target area of interest. Following administration, there is a period of time, known as the uptake period, during which the radiotracer circulates through the body and accumulates in the target area. This uptake period can vary depending on the radiotracer used and the biological process being studied, typically ranging from 30 minutes to several hours for conventional tracers, but potentially extending to days for the long-lived isotopes central to this invention. Once sufficient time has elapsed for optimal tracer distribution and target accumulation, a first image is acquired using a nuclear imaging technique such as positron emission tomography (PET) or single-photon emission computed tomography (SPECT). The choice between PET and SPECT depends on the isotope used and the specific diagnostic requirements. The first image captures the three-dimensional distribution of the radiotracer within the target area, providing an initial view of the region of interest. This image represents the raw data collected by the imaging system and may be of lower quality due to factors such as low signal-to-noise ratio, limited spatial resolution, or artifacts arising from the imaging process or patient movement.

[0031] The core of this invention lies in the integration of deep learning techniques with nuclear imaging to significantly improve image quality. This improvement in image quality can lead to earlier and more accurate detection and quantification of abnormalities within the target area. Early detection is crucial in medical diagnostics as it can lead to timely intervention and better patient outcomes. For certain applications, an accurate estimation of the tracer uptake is required, as the tracer uptake value may have prognostic value.

[0032] To enhance the quality of the first image, a deep learning model is employed. Deep learning models are a subset of machine learning algorithms that can learn and make decisions based on large datasets. The deep learning model used for image improvement is trained using a combination of supervised and unsupervised learning techniques. Supervised learning preferably involves using labeled datasets, where the input images are paired with high-quality reference images. Unsupervised learning, on the other hand, preferably involves techniques such as autoencoders, transformers, and generative adversarial networks (GANs) to learn the underlying structure of the images without explicit labels.

[0033] The use of deep learning models to improve image quality offers several advantages. By enhancing image quality, it is possible to use lower doses of radiotracers, thereby minimizing patient exposure to harmful radiation. This reduction in radiation dose is especially advantageous in medical imaging, where patient safety is of utmost importance. The ability to produce high-quality images with lower radiotracer doses can lead to more frequent and safer imaging procedures, allowing for better monitoring and diagnosis of medical conditions.

[0034] The use of a deep learning model to enhance image quality preferably allows for clearer and more defined differentiation between various tissues. This improved differentiation is particularly advantageous in diagnostic evaluations, as it enables more precise identification and analysis of abnormalities or conditions within the target area. For instance, in the context of oncology, the enhanced images may assist in better delineating tumor boundaries, thereby aiding in more accurate staging and treatment planning. In a further embodiment, the deep learning model may be continuously updated with new imaging data, ensuring that its performance improves over time. This continuous learning capability is particularly beneficial in adapting to new types of radiotracers and imaging conditions, thereby maintaining the accuracy and reliability of the enhanced images. Additionally, the deep learning model may incorporate advanced techniques such as transfer learning, where knowledge from one imaging task is transferred to another, further enhancing its versatility and effectiveness.

[0035] The method of nuclear imaging as described herein provides significant advantages in terms of improved image quality, which in turn facilitates more accurate and effective medical treatments. By leveraging advanced deep learning techniques, the method enhances the clarity and detail of nuclear images, making it a valuable tool in the field of medical imaging.

[0036] Preferably, the method can be integrated into existing nuclear imaging systems with minimal modifications. This integration is facilitated by the use of standardized image formats and interfaces, which allow for seamless communication between the imaging system and the deep learning model. Preferably, the integration process includes validation steps to ensure that the enhanced images meet clinical standards for diagnostic accuracy.

[0037] In an embodiment, said deep learning model comprises a convolutional neural network (CNN), preferably trained on a diverse dataset of nuclear images.

[0038] In a preferred embodiment, the method utilizes an ensemble 3D U-Net architecture.

[0039] In a more preferred embodiment, the method utilizes an ensemble 3D U-Net architecture, wherein each of the ensemble models are trained on differing noise levels. This approach contributes to more robust imaging, making the method more reliable under different conditions. The ensemble 3D U-Net architecture preferably involves at least two separate U-Net models, each designed to handle different aspects of image reconstruction.

[0040] Preferably, one U-Net model is trained to reduce high-frequency noise, which is common in initial nuclear imaging scans. This model may employ a variety of noise reduction techniques, including but not limited to, Gaussian filtering, median filtering, and wavelet transforms. The second U-Net model is preferably trained to enhance low-frequency details, ensuring that important anatomical and functional information is preserved. This model might utilize advanced techniques such as edge detection, contrast enhancement, and texture mapping.

[0041] More preferably, every U-Net is trained on a subset of training PET images exhibiting specific noise characteristics. The noise characteristics of the training images may be determined by applying one or more noise filters, such as but not limited to Gaussian, Perona-Malik, Rician, Poisson, or wavelet-based filters, to the training images and categorizing the images based on filter response values. Alternative or additional noise characterization methods may be employed. The outputs of the plurality of U-Net models may be combined using weighted fusion, wherein the weights assigned to each U-Net output may be determined dynamically or adaptively as a function of estimated noise characteristics of the input image to be enhanced, such as image filters, for example Gaussian, Perona-Malik, Rician, Poisson, or wavelet-based filters.

[0042] The training data for these models preferably includes a wide range of noise levels, from very low to very high, to ensure robustness. Different noise levels of the same image can be obtained, by reconstructing images with reduced count events data. The noise levels range from 1% to 100% of the clinical counts signal intensity, preferably from 5% to 75%, more preferably from 10% to 60%, and most preferably from 25% to 50%. This diverse training dataset ensures that the models can handle various imaging conditions, making the method highly adaptable.

[0043] Preferably, the method includes a preprocessing step where the first image undergoes normalization, artifact removal and alignment to ensure consistency before being input into the deep learning model. This preprocessing step may also involve the application of filters to remove artifacts that could interfere with the deep learning model's performance. Preferably, the preprocessing techniques are selected based on the characteristics of the target area and the type of radiotracer used.

[0044] In an embodiment, the isotope-labeled radiotracer was not included in the training dataset of the deep learning model.

[0045] In an embodiment, the training dataset consists preferably partially, and more preferably exclusively, of images obtained using short-lived radiotracers, such as 18F-labeled tracers (e.g., [18F]FDG, [18F]PSMA, [18F]FLT) and 68Ga-labeled tracers (e.g., [68Ga]PSMA, [68Ga]DOTATATE). Despite being trained on these short- lived tracer images, the model is designed to be applicable to images obtained with long-lived tracers. The training process involves feeding the model pairs of low- quality and corresponding high-quality images from these short-lived tracer scans, enabling it to learn the complex transformations needed to enhance image quality. The model learns to recognize patterns and features that are indicative of high- quality images, such as clear organ boundaries, distinct uptake regions, and appropriate contrast levels. When applying the trained deep learning model to the first image obtained with long-lived tracers, it performs several key functions:

[0046] • Noise reduction: The model identifies and suppresses random fluctuations in image intensity that do not represent true signal.

[0047] • Artifact removal : Common artifacts, such as streak artifacts in PET images, are recognized and mitigated.

[0048] • Resolution enhancement: The model applies learned patterns to increase the apparent resolution of the image, making fine structures more visible.

[0049] • Contrast improvement: The difference between areas of high and low tracer uptake is accentuated, improving overall image clarity.

[0050] The deep learning model processes the entire 3D volume of the first image, considering spatial context in all three dimensions. This approach allows for more sophisticated enhancement compared to traditional 2D image processing techniques. The output of this deep learning enhancement is the second image, which maintains the true signal from the original scan while significantly improving overall image quality. The second image provides a clearer and more detailed view of the target area. This enhanced image can reveal abnormalities that may not be visible in the first image, thereby aiding in the early detection of diseases such as cancer, cardiovascular disorders, and neurological conditions. In a further embodiment, the deep learning model can be continually updated and refined with new data to improve its performance over time. This iterative learning process ensures that the model remains effective in enhancing image quality as new imaging techniques and radiotracers are developed. Preferably, the deep learning model enhances the image resolution, allowing for the identification of finer structural details within the target area. This enhanced resolution is particularly advantageous for medical diagnostics, where the ability to discern small anatomical structures can be critical for accurate diagnosis and treatment planning. In a more preferred embodiment, the deep learning model may incorporate various image processing techniques such as noise reduction, contrast enhancement, and edge detection to further improve the image quality. These techniques can be applied individually or in combination, depending on the specific requirements of the imaging task. Additionally, the deep learning model is preferably trained on a large and diverse dataset of nuclear images to ensure robust performance across a wide range of imaging conditions. This comprehensive dataset is designed to capture the variability encountered in real-world clinical scenarios. More preferably, the training dataset includes images with varying levels of noise and resolution to make the model more adaptable to different imaging scenarios. The diversity of the training dataset extends to several key factors:

[0051] • Radiotracer doses: The dataset includes images acquired using different radiotracer doses, ranging from low-dose protocols to standard and high- dose acquisitions. This variation allows the model to adapt to images obtained under various radiation exposure levels, enhancing its ability to improve image quality across different dosing protocols.

[0052] • Acquisition durations: Images with different acquisition durations are included, from rapid scans to extended imaging sessions. This variety enables the model to handle temporal variations in image quality and noise levels associated with different scan times.

[0053] • Patient diversity: The training data encompasses images from a diverse patient population, including variations in age, gender, body mass index, and ethnicity. This broad representation helps the model generalize across different patient characteristics and anatomical variations.

[0054] • Scanner variety: Images acquired from various PET and SPECT scanner brands and models are incorporated into the dataset. This inclusion allows the model to adapt to subtle differences in image characteristics that may arise from different scanner technologies and reconstruction algorithms.

[0055] • Reconstruction techniques: The dataset includes images reconstructed using various algorithms and parameter settings, enabling the model to handle differences in image texture and quality arising from different reconstruction techniques.

[0056] • Disease states and tracer distribution patterns: Images representing various disease states and tracer distribution patterns are included, allowing the model to enhance images across different clinical scenarios and uptake patterns.

[0057] • Artifacts: The dataset incorporates images with common artifacts such as motion artifacts, attenuation correction errors, and scatter, training the model to recognize and potentially correct for these issues.

[0058] By training on such a comprehensive and varied dataset, the deep learning model becomes more robust and versatile, capable of enhancing image quality across a wide spectrum of clinical scenarios, imaging protocols, and technical variations. This extensive training approach ensures that the model can generalize well to new, unseen data, making it a reliable tool for improving image quality in diverse nuclear imaging applications.

[0059] In an embodiment, the method may also include a step of validating the second image against a ground truth image to assess the accuracy of the image enhancement. This validation step can be performed using various metrics such as peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and root mean square error (RMSE), or others, or any combination of these.

[0060] While many common PET radiotracers such as 18F and 68Ga have relatively short half-lives (110 minutes and 68 minutes, respectively), this invention focuses on the use of isotope-labeled radiotracers with substantially longer half-lives. In an embodiment, said isotope is chosen from the group consisting of 89Zr, 64CU, 1241, 52Mn, and 55Co. More preferably, said isotope is chosen from 89Zr and 1241. The method as described herein uses radiotracers with a half-life of at least 750 minutes (12.5 hours), which ensures that the radiotracer remains active for a sufficient duration to acquire high-quality images over extended periods, as claimed in claim 1. The half-life of the isotope is a critical parameter, and in various embodiments, the half-life can range from 750 minutes to 9000 minutes (150 hours, about 6.25 days). Preferably, the half-life is between 900 minutes (15 hours) and 5400 minutes (90 hours), more preferably between 1080 minutes (18 hours) and 4800 minutes (80 hours), even more preferably between 1200 minutes (20 hours) and 3600 minutes (60 hours), and most preferably between 1440 minutes (24 hours) and 2880 minutes (48 hours). These ranges provide an optimal balance between extended image acquisition time and practical considerations, while ensuring sufficient radioactivity for high-quality imaging. Examples of isotopes with half-lives within these ranges include 64Cu (half-life: 12.7 hours), 55Co (half-life: 17.5 hours), 89Zr (half-life: 78.4 hours), 1241 (half-life: 100.2 hours), and 52Mn (halflife: 134.2 hours). The use of these longer-lived isotopes allows for extended imaging sessions, which can be beneficial for studying slow biological processes, capturing the full kinetics of drug distribution, or performing longitudinal studies over several days. Moreover, some of these long-lived radiotracers are used to bind cells, peptides, proteins or antibodies.

[0061] In an embodiment, said first image is a low-grade quality image. A low-grade quality image is an image with a signal-to-noise ratio below 3, where the signal-to-noise ratio is calculated by dividing the mean standard uptake value (SUVmean) by the standard deviation in standard uptake value, calculated in a uniform, spheric region of interest (ROI) in the liver with a diameter of at least 3 centimeters. Preferably, the signal-to-noise ratio in the same ROI in the second image is above 3, more preferably above 5, and most preferably above 8.

[0062] In a preferred embodiment, said nuclear image scan is a PET scan or a SPECT scan.

[0063] In an embodiment, said radiotracer is an isotope-labelled biological agent, such as a cell (e.g., a T cell), a peptide, a protein, or an antibody. Most commonly the radiotracers are labelled to a specific antibody or nanobody, such that the slow pharmacokinetics of these labelled agents can be imaged Therefore, in a preferred embodiment, said biological agent is an antibody, preferably a nanobody (VHH fragment) or a monoclonal antibody.

[0064] In a further embodiment, said antibody is able to bind to T-cells, B-cells, or NK-cells, or is a therapeutic agent, such as but not limited to Alemtuzumab, Bevacizumab, Cetuximab, Crefmirlimab, Daclizumab, Dostarlimab, Durvalumab, Erenumab, Gemtuzumab ozogamicin, Ibritumomab tiuxetan, Ipilimumab, Nivolumab, Ofatumumab, Panitumumab, Pembrolizumab, Ranibizumab, Rituximab, Sacituzumab govicetan, or Trastuzumab.

[0065] Crefmirlimab, specifically known as 89Zr crefmirlimab berdoxam, is an investigational radiotracer used in immunoPET imaging. It is designed to target CD8 receptors on human T cells, which are crucial components of the immune response, particularly in cancer immunotherapies. By binding to these CD8 receptors, crefmirlimab allows for non-invasive PET imaging to visualize and quantify the distribution of CD8+ T cells in the body. This capability is particularly useful for assessing the immune status of patients, measuring the effectiveness of immunotherapies, and potentially predicting patient outcomes. Crefmirlimab is labeled with the radioactive isotope zirconium-89, making it detectable via PET scans. It is being investigated in clinical trials to support its use alongside various immunotherapies, such as CAR-T cell treatments for solid tumors. Durvalumab is a monoclonal antibody that targets the PD-L1 protein (Programmed Death-Ligand 1) on cancer cells. By binding to PD-L1, durvalumab blocks its interaction with the PD- 1 receptor on T cells, thereby preventing the "off" signal that cancer cells use to evade the immune system. This allows the immune system to better detect and attack cancer cells. Durvalumab is commonly used to treat certain types of cancers, including non-small cell lung cancer (NSCLC) and urothelial carcinoma (a type of bladder cancer). It is often used as a maintenance therapy after initial chemotherapy has reduced the size of the tumor, helping to prolong the period during which the cancer remains under control.

[0066] Here, the biodistribution of such antibodies would be indicative of therapy response, and disease staging or progression, especially for most of the currently explored targets. This approach provides flexibility in selecting the appropriate isotope based on the specific imaging requirements, target area, and biological processes under investigation, offering advantages over conventional short-lived radiotracers in certain applications.

[0067] In a second aspect, the current invention also relates to a computer-implemented method for analyzing and / or improving a nuclear image, said method comprises receiving a first nuclear image, wherein said first image is acquired using an isotopelabeled radiotracer, wherein said isotope has a half-life of at least 750 minutes, and obtaining a second image from said first image, wherein said second image is obtained by improving said first image quality by means of a deep learning model.

[0068] It will be clear to a skilled person that the embodiments as discussed above are also applicable to this computer-implemented method. In an embodiment, the deep learning model is an ensemble 3D U-Net architecture. In a preferred embodiment, said combined model output is obtained by adaptively weighting the separate U-Net model outputs.

[0069] In an embodiment the adaptive weighting coefficients are determined based on the output obtained by applying a computer vision filter to either the input image, the output images of each respective model or any combination of the input image and output images of each model.

[0070] In a further aspect, the current invention also relates to a method of diagnosing a disorder or disease. Said method includes a nuclear imaging step as in the embodiments explained above. Said method is used for the diagnosis or detecting a tumoral growth, such as breast cancer, biliary tract cancer, endometrial cancer, hepatocellular cancer, non-small cell lung cancer, kidney cancer, skin cancer and small cell lung cancer.

[0071] In a final aspect, the current invention also relates to a method for denoising positron emission tomography (PET) images using deep learning, comprising the steps of: a) training a deep learning model on PET images; b) applying the trained model to PET images of novel radiotracers, previously unseen by the model, to enhance image quality.

[0072] In a preferred embodiment, the deep learning model is an ensemble 3D U-Net architecture, wherein the ensemble 3D U-Net architecture comprises at least two models trained on datasets characterized by differing levels of noise.

[0073] In a further embodiment, said ensemble U-Net outputs are dynamically weighted as a function of the input image.

[0074] In an embodiment, the deep learning model is trained using PET images reconstructed from list-mode raw data at reduced radioactive event counts.

[0075] Preferably, the PET images used for training are those of [68Ga]Ga-PSMA, [68Ga]Ga-DOTATATE and [18F]FDG.

[0076] In an embodiment, said images are obtained by means of long-lived PET tracers, such as [89Zr]Zr-Df-crefmirlimab and [89Zr]Zr-DFO-durvalumab.

[0077] Disclosed herein are also computer program products or a non-transitory computer- readable medium comprising instructions stored on a non-transitory computer- readable medium, the instructions being configured, when executed by one or more processors, to cause the one or more processors to: receive a first nuclear image acquired using an isotope-labeled radiotracer having a half-life of at least 750 minutes; process the first nuclear image with a deep learning model, wherein the deep learning model is an ensemble 3D U-Net architecture comprising at least two models trained on datasets with differing noise characteristics; and generate a second nuclear image by enhancing the quality of the first nuclear image based on dynamically weighted outputs of the deep learning model.

[0078] In an embodiment, said computer program comprises instructions which, when executed by one or more processors, cause the one or more processors to perform the methods as described herein. The computer program may be stored on a non-transitory computer-readable medium, such as a hard disk, optical disk, flash memory, or cloud-based storage system. Execution of the program enables the automatic enhancement of nuclear images acquired with long-lived isotope-labeled radiotracers by applying a deep learning model, such as an ensemble 3D U-Net architecture trained on datasets with varying noise characteristics.

[0079] In yet another aspect, the invention provides a system for nuclear image enhancement. The system comprises at least one processor and at least one memory storing a computer program as described above. The processor is configured to receive a first nuclear image obtained using an isotope-labeled radiotracer, execute the computer program to process the first image by means of the deep learning model, and output a second image of enhanced quality. The system may further include an input interface for receiving nuclear images from a PET or SPECT scanner, and an output interface for transmitting enhanced images to a display or clinical database. Preferably, the system is integrated into, or operable in connection with, existing nuclear imaging devices, thereby enabling real-time or near real-time image enhancement during clinical workflows.

[0080] The invention will now be described by means of figures and examples that are not limitative to the current invention.

[0081] FIGURES

[0082] Figure 1 illustrates a schematic overview of an embodiment of the invention in which low-grade nuclear images are enhanced through deep learning. The low-grade quality images serve as input to an ensemble of U-Net models (UNet 1 ... UNet N), each trained on image datasets with different noise characteristics. Each model generates a corresponding set of output images (Output Images 1 ... Output Images N). These outputs are then combined using a dynamic output weighting process, which adaptively assigns weights based on the characteristics of the input image or the intermediate model outputs. The resulting combined output yields enhanced images with improved resolution, contrast, and signal-to-noise ratio compared to the original low-grade images.

[0083] Figure 2 illustrates a schematic overview of an embodiment of the invention in which a low-grade quality nuclear image, acquired using a long-lived isotope-labeled radiotracer, is enhanced by means of a deep learning model. The model has been trained on datasets obtained from short-lived radiotracers, yet it is capable of generalizing to unseen long-lived tracer data. The figure shows the process wherein the low-quality input image is processed by the trained deep learning model, resulting in an enhanced image with improved resolution, reduced noise, and superior diagnostic quality compared to the original image.

[0084] Figures 3 and 4 are discussed in the example section.

[0085] EXAMPLES

[0086] Example 1: Enhancing Zr-89 and Cu-64 PET Phantom Image Quality using Deep Learning Denoising for Novel Radiotracers

[0087] Aim / Introduction

[0088] The FDA-approved radiopharmaceutical 68Ga-DOTATATE for PET imaging offers high sensitivity and specificity for detecting Suspected Somatostatin Receptor-Positive (SSTR)-expressing neuroendocrine tumors (NETs) but is limited by its short half-life and production constraints. In contrast, 64Cu-DOTATATE, an investigational PET radiotracer, demonstrates a lower radiation dose, superior lesion detection rates, improved spatial resolution, and a longer half-life, which enhances its practicality for routine clinical use. These attributes make 64Cu-DOTATATE a promising alternative for SSTR-based imaging in patients with NETs. Despite this, the accurate quantification of 64Cu-DOTATATE remains challenging and warrants sufficiently long scan times. The aim of this study is to assess if the scan time of 64Cu-DOTATATE can be improved, without compromising on the accuracy of SUV quantification. Furthermore, the analysis extends to studies involving unchelated 89Zr, a long-lived isotope increasingly utilized in clinical practice for labeling monoclonal antibodies, thereby facilitating the diagnosis, monitoring, and treatment of various oncologic diseases.

[0089] Materials and Methods

[0090] First, a deep-learning model with a U-Net architecture was trained on matching low- count and standard-count PET scans from two different hospitals. During training, only matching pairs of two commonly available tracers were seen by the model, [68Ga]Ga-PSMA and [18F]FDG. The model was then validated on PET scans of the IEC NEMA phantom filled with 64Cu-DOTATATE and an unchelated 89Zr solution. Scans were reconstructed at different acquisition times, ranging from 60s to 1200s, by downsampling counts of the list-mode data. Image quality was assessed using the coefficient of variation in relevant regions of interest (ROIs) and through visual inspection comparing the original scans with their enhanced counterparts.

[0091] Results

[0092] Validation on the IEC NEMA phantom confirmed that the matching enhanced images had a lower variance in SUV, as shown in Figure 3, across all spheres and scan times. The enhanced method consistently reduced SUV variance, particularly in smaller spheres and shorter scan durations, where the improvement is most pronounced. Additionally, the recovery coefficient for SUVmax improved in all spheres, as demonstrated in Figure 4. The enhanced reconstruction method not only increased the RC SUVmax but also brought it closer to the desired value of 1, particularly for the smaller spheres at shorter scan durations (300s and 600s). This indicates that the enhancement effectively compensates for the lower count statistics typically associated with shorter scans, leading to more accurate quantification.

[0093] Conclusion

[0094] In silico denoising by our deep-learning-based model seems to at least partially generalize to 64Cu- and 89Zr-labelled radiotracers as shown by this study on phantom studies acquired with 64Cu-DOTATATE and 89Zr. Strategies for post-hoc enhancement of PET-scans allows for a more confident image assessment and may help to harmonize scan quality, which facilitates pooled analyses of multi-center studies using less commonly used radiotracers.

[0095] Example 2: Assessing the Quantification Accuracy of Al Enhanced Cu-64 and 89-Zr Labelled Radiotracer Clinical PET Images

[0096] Aim / Introduction

[0097] While [18F]FDG has long been the reliable workhorse of nuclear imaging, there is an increasing need for more sensitive and specific radiotracers to improve diagnosis and treatment planning. An interesting avenue is the use of Zr-89 labelled monoclonal antibodies, as the longer half-time of Zr-89 matches the slow pharmacokinetics of these antibodies. Similarly, 64Cu-DOTATATE offers promising sensitivity and specificity rates for detecting Suspected Somatostatin Receptor- Positive (SSTR)-expressing neuroendocrine tumors (NETs). However, such less established tracers typically suffer from high costs and varying image quality, impeding broader adaptation. The use of deep-learning based denoising algorithms could help remediate these challenges, but generally require many scans for training, which are not commonly available for new tracers. This study aims to evaluate the generalizability of deep learning-based denoising algorithms, trained on conventional, short-lived radiotracers, to less common tracer types and radioisotopes that were not encountered during model training. More specifically, clinical PET / CT scans acquired with the novel radiotracers 64Cu-DOTATATE, [89Zr]Zr-DFO-durvalumab and [89Zr]Zr-Df-crefmirlimab are used as validation data.

[0098] Materials and Methods

[0099] First, a deep-learning model with a U-Net architecture was trained on matching low- count and standard-count PET scans from two different hospitals. During training, only matching pairs of two commonly available tracers were seen by the model, [68Ga]Ga-PSMA and [18F]FDG. In total, twelve patient scans were acquired, each using one of the following tracers: 64Cu-DOTATATE, [89Zr]Zr-DFO-durvalumab, or [89Zr]Zr-Df-crefmirlimab. The scans were acquired in list-mode reconstructed at 100%, 50%, and 25% of the count levels used in clinical practice. All lesions were annotated on the 100% scans by a certified nuclear physician. Then, the 50% and 25% scans were enhanced using the deep-learning model and the quantification of SUV was compared between the clinical and enhanced images.

[0100] Results

[0101] The deep learning model could significantly improve image quality of the 64Cu- DOTATATE, [89Zr]Zr-DFO-durvalumab and the [89Zr]Zr-Df-crefmirlimab scans, strongly reducing the variation in SUV, which is a well-known indicator of PET noise, for several organs across different patients. The algorithm also removed some unrealistically high SUVmax that were observed in certain tissues and overall resulted in a visually more appealing scan.

[0102] In total, 23 lesions were annotated, ranging from 9 to 107 voxels in size. In general a strong correlation could be observed between the SUVmean on the clinical and the 50% enhanced (Spearman p=0.91) and between the clinical and the 100% enhanced scans (Spearman p=0.87). In absolute terms the quantifications on the 50%-enhanced scans were closer to the 100% scans as quantified by the RMSE, compared to the 25%-enhanced scans. Nonetheless, the 25%-enhanced scans scored better than both the 25%- and 50%-reduced scans, highlighting a clear improvement in quantification by the Al algorithm. A more detailed Bland Altman analysis revealed a slight bias among the 25%-enhanced images for smaller lesions. For smaller lesions a systematic underestimation of the SUV could be observed. For the 50%-enhanced images, no such bias could be observed. Finally, we also assessed image quality as defined by the signal to noise ratio (SNR) measured on spherical ROI in the liver. The SNR of the 50%-enhanced image was found to be slightly higher than the clinical image and significantly higher than the 25%- enhanced image. In terms of SNR as well, the 25%-enhanced image outperformed both the 50%- and 25%-reduced scans.

[0103] Conclusion

[0104] In this work, we have demonstrated that a deep learning model, trained on commonly available radiotracer data allows to accelerate 64Cu-DOTATATE, [89Zr]Zr-Df-crefmirlimab, and [89Zr]Zr-DFO-durvalumab PET scans, without compromising on quantification accuracy. Such accelerated scanning may improve patient comfort or could enable the acquisition of PET scans with long-lived radiotracers with lower doses of injected radiotracer.

Claims

CLAIMS1. A method of nuclear imaging, wherein said method uses of an isotopelabeled radiotracer, wherein said isotope has a half-life of at least 750 minutes, said method comprises acquiring a first image by means of a nuclear imaging method of a target area, and obtaining a second image from said first image, wherein said second image is obtained by improving said first image quality by means of a deep learning model.

2. The method according to claim 1, wherein the deep learning model comprises a convolutional neural network (CNN).

3. Method according to claim 2, wherein the deep learning model is an ensemble 3D U-Net architecture.

4. Method according to claim 3, wherein the ensemble 3D U-Net architecture comprises at least two models trained on datasets characterized by differing levels of noise.

5. Method according to claim 1, wherein the isotope-labeled radiotracer was not included in the training dataset of the deep learning model.

6. Method according to any of the previous claims, wherein said isotope is chosen from the group consisting of 89Zr, 64CU, 1241, 52Mn, and 55Co.

7. Method according to any previous claims, wherein said first image is a low- grade quality image.

8. Method according to any of the previous claims, wherein said nuclear image scan is a PET scan or a SPECT scan.

9. Method according to any of the previous claims, wherein said radiotracer is an isotope-labelled biological agent, such as a cell (e.g., a T cell), a peptide, a protein, or an antibody.

10. Method according to claim 9, wherein said biological agent is an antibody, preferably a nanobody (VHH fragment).

11. Method according to claim 10, wherein said antibody is able to bind to T- cells, B-cells, or NK-cells, or is a therapeutic agent, such as Alemtuzumab,Bevacizumab, Cetuximab, Crefmirlimab, Daclizumab, Dostarlimab, Durvalumab, Erenumab, Gemtuzumab ozogamicin, Ibritumomab tiuxetan, Ipilimumab, Nivolumab, Ofatumumab, Panitumumab, Pembrolizumab, Ranibizumab, Rituximab, Sacituzumab govicetan, or Trastuzumab.

12. A computer-implemented method for analyzing and / or improving a nuclear image, said method comprises receiving a first nuclear image, wherein said first image is acquired using an isotope-labeled radiotracer, wherein said isotope has a half-life of at least 750 minutes, and obtaining a second image from said first image, wherein said second image is obtained by improving said first image quality by means of a deep learning model.

13. The computer-implemented method according to claim 12, wherein the deep learning model is an ensemble 3D U-Net architecture.

14. The computer-implemented method according to claim 13, wherein the ensemble 3D U-Net architecture comprises at least two models trained on datasets characterized by differing levels of noise.

15. The computer-implemented method according to claim 14, wherein the combined model output is obtained by adaptively weighting the separate U- Net model outputs.

16. The computer-implemented method according to claim 15, wherein the adaptive weighting coefficients are determined based on the output obtained by applying a computer vision filter to either the input image, the output images of each respective model or any combination of the input image and output images of each model.

17. A method of diagnosing a disorder or disease, said method includes an nuclear imaging step according to any of the claims 1 to 16.

18. The method according to claim 17, wherein said method is used for the diagnosis or detecting a tumoral growth, such as breast cancer, biliary tract cancer, endometrial cancer, hepatocellular cancer, non-small cell lung cancer, kidney cancer, skin cancer and small cell lung cancer.

19. A method for denoising positron emission tomography (PET) images using deep learning, comprising the steps of: a) training a deep learning model onPET images; b) applying the trained model to PET images of novel radiotracers, previously unseen by the model, to enhance image quality.

20. The method of claim 19, wherein the deep learning model is an ensemble 3D U-Net architecture, wherein the ensemble 3D U-Net architecture comprises at least two models trained on datasets characterized by differing levels of noise.

21. The method of claim 20, wherein the ensemble U-Net outputs are dynamically weighted as a function of the input image.

22. The method of claim 21, wherein the deep learning model is trained using PET images reconstructed from list-mode raw data at reduced radioactive event counts.

23. The method of any of the claims 19 to 22, wherein the PET images used for training are those of [68Ga]Ga-PSMA, [68Ga]Ga-DOTATATE and [18F]FDG.

24. The method of claim 19, wherein said images are obtained by means of long- lived PET tracers, such as [89Zr]Zr-Df-crefmirlimab and [89Zr]Zr-DFO- durvalumab.

25. A computer program comprising instructions which, when executed by one or more processors, cause the one or more processors to perform a method according to any of claims 1 to 24.

Images