Method and system for normalizing nuclear imaging data using deep learning

Abstract

The invention provides a method for normalizing nuclear imaging data by processing it through a deep learning algorithm, specifically a neural network comprising at least one convolutional layer, to reduce inter-center, intra-center, and patient weight biases. The method accepts data from diverse sources, including different imaging centers, scanner models, and patient populations, and applies bias reduction techniques, such as normalizing image intensity, to produce normalized data. This normalized data is then utilized in various downstream tasks, such as computer-aided diagnosis and clinical trials, yielding improved accuracy compared to non-normalized data. The neural network is designed to adapt to new data sources without retraining, enhancing the method's flexibility.

Description

[0001] METHOD AND SYSTEM FOR NORMALIZING NUCLEAR IMAGING DATA

[0002] USING DEEP LEARNING

[0003] FIELD OF THE INVENTION

[0004] The present invention relates to methods for normalizing nuclear imaging data, specifically addressing the reduction of inter-center, intra-center, and patient weight biases in nuclear imaging data obtained from multiple sources. The invention utilizes deep learning algorithms, including convolutional neural networks (CNNs), to process and normalize data from various imaging origins such as different imaging centers, scanner hardware models, reconstruction algorithms, acquisition protocols, radiotracer dosing schemes, imaging time points post-radiotracer administration, and patient population characteristics. The normalized data can be used in downstream analysis tasks, including computer-aided diagnosis, denoising, autoencoding, classification, prognosis, disease classification, treatment response prediction, biomarker detection, image segmentation, radiomics feature extraction, longitudinal patient monitoring, drug development assessment, clinical trial analysis, and epidemiological studies, thereby improving the accuracy of these tasks compared to using non-normalized data.

[0005] BACKGROUND

[0006] Nuclear imaging techniques, such as positron emission tomography (PET) and single-photon emission computed tomography (SPECT), are critical tools in medical diagnostics and research. These techniques provide detailed images of metabolic processes and other physiological functions within the body. However, the accuracy and reliability of nuclear imaging data can be significantly affected by various sources of bias. These biases can arise from differences in imaging centers, scanner hardware models or manufacturers, reconstruction algorithms, acquisition protocols, radiotracer dosing schemes, imaging time points post-radiotracer administration, and patient weight. Inter-center, intra-center, and patient weight biases can lead to inconsistencies in the quantification of tracer uptake on the imaging data, which in turn can negatively impact the outcomes of clinical trials, disease diagnosis, treatment response prediction, and other critical medical and research applications. Current methods to address these biases are often inadequate, leading to challenges in achieving reliable and comparable nuclear imaging data across different sources. This lack of normalization hampers the ability to conduct multi-center clinical trials and other collaborative research efforts, ultimately impacting the accuracy and efficacy of medical diagnostics and treatments.

[0007] SUMMARY OF THE INVENTION

[0008] The invention relates to a method for normalizing nuclear imaging data, which involves receiving at least two different nuclear imaging scans from one or more different patients processing it using a deep learning algorithm to reduce intercenter, intra-center, and patient weight biases. A normalized scan is then generated with a reduced bias. The sources of variation in nuclear imaging data can vary widely, including different imaging centers, scanner hardware models, reconstruction algorithms, acquisition protocols, radiotracer dosing schemes, imaging time points post-radiotracer administration, and patient population characteristics. The method is applicable to data from positron emission tomography (PET) and single-photon emission computed tomography (SPECT). Bias reduction includes normalizing image intensity across sources and can be quantified to ensure effectiveness. The normalized data is useful for multi-center clinical trials and various downstream analysis tasks such as computer-aided diagnosis, disease classification, treatment response prediction, and more, yielding improved accuracy compared to non-normalized data. The method also includes a computer- implemented version that uses a deep learning model trained on high-signal, metabolically relevant regions. The deep learning algorithm includes an encoder network fine-tuned to reduce biases while preserving task-specific information. This method adapts to new imaging data sources without retraining, ensuring consistency and reliability in nuclear imaging data analysis. The advantages of this invention include reduced variability, improved diagnostic reliability, faster processing, enhanced longitudinal patient monitoring, and better consistency in clinical trial outcomes.

[0009] In various aspects, the present invention may be embodied as a method, a system, or a non-transitory computer-readable medium. In the method aspect, the invention provides a sequence of steps as described herein, including receiving nuclear imaging data from one or more sources, applying a deep learning algorithm to mitigate inter-center, intra-center, and patient weight biases, and generating normalized nuclear imaging data with reduced biases. In the system aspect, the invention comprises one or more processors and a memory storing instructions that, when executed, cause the one or more processors to carry out the method as described. In the computer-readable medium aspect, the invention provides a tangible storage medium containing program instructions that, when executed by one or more processors, likewise perform the method as described. In preferred embodiments, the system and medium embodiments further support additional features such as quantifying the reduction in biases, adapting to new sources of nuclear imaging data without retraining, and outputting normalized data for downstream applications including computer-aided diagnosis, disease classification, treatment response prediction, biomarker detection, longitudinal patient monitoring, and multi-center clinical trials.

[0010] DESCRIPTION OF FIGURES

[0011] Figure 1 shows a schematic overview of a possible embodiment of a method for normalizing input images originating from different imaging centers and scanning protocols with deep learning. In this embodiment, a trained classifier assesses the performance of the normalizer model by performing a tumor triaging task.

[0012] Figure 2 shows a scatter plot of standardized uptake value (SUVmax) as a function of weight for both original (grey circles) and harmonized (black crosses) nuclear imaging data.

[0013] Figure 3 and 4 are described in the example sections.

[0014] DETAILED DESCRIPTION OF THE INVENTION

[0015] The current invention relates to methods of nuclear imaging employing deep learning models to reduce bias due to non-physiological factors.

[0016] Definitions

[0017] The term "nuclear imaging data" refers in the present invention to data obtained from imaging techniques that utilize radioactive substances to visualize and measure biological processes in the body. This includes, but is not limited to, positron emission tomography (PET) and single-photon emission computed tomography (SPECT).

[0018] By the term "multiple sources" is meant in the present invention at least two different imaging origins. These origins can include different imaging centers, different scanner hardware models or manufacturers, different reconstruction algorithms, different acquisition protocols, different radiotracer dosing schemes, different imaging time points post-radiotracer administration, and patient weight.

[0019] The term "deep learning algorithm" refers in the present invention to a type of machine learning algorithm that uses multiple layers of artificial neural networks to model complex patterns in data. Specifically, this includes convolutional neural network (CNN) layers and other architectures suitable for processing imaging data.

[0020] The term "standardized uptake value (SUV)" refers to patient-weighted and decay- corrected measurements of the activity concentration in a region of interest within PET imaging. It quantifies the uptake of radiotracer in tissues, normalized by the injected dose and the patient's body weight, thereby allowing for standardized comparison across different patients and imaging sessions.

[0021] By the term "reducing inter-center, intra-center, or patient weight biases" is meant in the present invention the process of minimizing non-physiological variations in the imaging data that arise due to differences between imaging centers, within the same imaging center over time, or patient weight. This can involve techniques such as normalizing image intensity across the multiple sources.

[0022] The term "normalized nuclear imaging data" refers in the present invention to nuclear imaging data that has been processed to reduce biases and variations, resulting in data that is more consistent and comparable across different sources.

[0023] By the term "quantifying the reduction in biases" is meant in the present invention the process of measuring the extent to which biases have been minimized in the normalized nuclear imaging data. This can involve statistical analysis and validation protocols.

[0024] The term "multi-center clinical trials" refers in the present invention to clinical studies conducted across multiple research centers or hospitals, where normalized nuclear imaging data is used to ensure consistency and reliability of imaging results.

[0025] By the term "downstream analysis task" is meant in the present invention any subsequent processing or analysis performed on the normalized nuclear imaging data. This includes tasks such as computer-aided diagnosis, computer-aided prognosis, disease classification, treatment response prediction, biomarker detection, image segmentation, radiomics feature extraction, longitudinal patient monitoring, drug development assessment, clinical trial analysis, and epidemiological studies.

[0026] The term "convolutional neural network (CNN)" refers in the present invention to a type of deep learning algorithm characterized by the use of convolutional layers, which are particularly effective for processing grid-like data such as images. CNNs are trained by sampling 3D volumes of voxels from high-signal, metabolically relevant regions in the nuclear imaging data.

[0027] By the term "encoder network" is meant in the present invention a part of a neural network that is trained to generate encoded embeddings of the nuclear imaging data. This network captures relevant features for normalization while preserving task-specific information from an initial classification task. The encoder network is fine-tuned to reduce biases in the encoded embeddings, and the normalized nuclear imaging data is reconstructed from these bias-reduced embeddings.

[0028] The term "adapt to new sources of nuclear imaging data without retraining" refers in the present invention to the capability of the deep learning algorithm to process and normalize data from new imaging sources without the need for additional training, thereby maintaining its performance and accuracy.

[0029] Detailed description

[0030] Described herein is a method for normalizing two or more nuclear imaging scans from one or more patients, comprising: receiving at least two nuclear imaging scans; applying a deep learning algorithm to process the scans and mitigate biases arising from variations in at least one of the following: acquisition protocols, reconstruction algorithms, reconstruction parameters, and patient weight; and generating normalized nuclear imaging data with reduced biases.

[0031] In a first aspect, the invention provides a method for normalizing nuclear imaging data, which includes receiving at least two nuclear imaging scans from one or more patients, processing the received nuclear imaging data using a deep learning algorithm, mitigating biases arising from variations in acquisition protocols, reconstruction algorithms, reconstruction parameters, and patient weight, generating normalized nuclear imaging data with reduced biases, and outputting normalized nuclear imaging data with reduced biases. Bias can be quantified using a correlation coefficient for continuous variables and using any statistical test such as the hypergeometric test, chi-squared test or Mann-Whitney U test between different categorical groups. Bias can be quantified using the resulting p-value of any such statistical procedure, with a lower p-value indicating a stronger bias. Using the proposed normalization techniques, this bias can be mitigated, which is reflected in a lower correlation coefficient and p-values when testing for non-physiological differences between these groups. To achieve a meaningful reduction in bias, the associated p-value should increase with preferably 5%, more preferably 10% and most preferably 20%.

[0032] This method ensures that the nuclear imaging data is consistent and reliable, thereby improving the accuracy of diagnostic and treatment decisions.

[0033] The primary advantage of this method is that it addresses the variability introduced by different imaging origins, such as different imaging centers, scanner hardware models, reconstruction algorithms, acquisition protocols, radiotracer dosing schemes, imaging time points post-radiotracer administration, and patient weight. By normalizing the nuclear imaging data, the method ensures that the data is comparable across different sources, which is crucial for multi-center clinical trials and other collaborative research efforts.

[0034] The method includes normalizing image intensity across multiple imaging sources to mitigate biases. This normalization process ensures that the images are comparable in terms of intensity values, which is crucial for accurate interpretation and analysis. Currently, nuclear imaging data rely on standardized values, corrected for decay time and patient weight, to facilitate comparisons between scans from different patients or the same patient at different points in time.

[0035] The normalization algorithm, trained on a diverse dataset from various scanners, has been optimized to identify and correct for technological artifacts specific to different scanner types that can produce physiologically implausible contrast recovery coefficients or SUVmax values. In current clinical practice, a deviation of up to 10% from the ground truth on phantom data is considered acceptable. Given that one scanning protocol might overestimate activity by 10%, while another might underestimate it by the same margin, the cumulative clinically permissible discrepancy could reach 20%. In such cases, the normalization software can reduce inter-technology differences by preferably 10%, more preferably 20%, and most preferably 30%, thereby enhancing the accuracy and reliability of nuclear imaging across different platforms. In an embodiment, the nuclear imaging data stems from at least two different imaging origins. With the origin of the image, several aspects of the imaging procedure can be implied. These aspects include variations in the hardware used, such as different scanner models or manufacturers, each of which may have unique sensitivity, resolution, and calibration characteristics that can impact the resulting images. Additionally, the reconstruction algorithms employed by different centers may vary, leading to differences in image quality and quantitative measures. These algorithms can affect how raw data is converted into the final image, introducing center-specific biases.

[0036] Moreover, imaging protocols, including acquisition settings such as the duration of the scan, the specific time points at which the imaging is performed post-radiotracer administration, and the radiotracer dosing schemes, can all introduce variability. For instance, images taken at different time points may show different tracer distributions due to biological processes, and different dosing schemes might affect the signal intensity, influencing the quantification of uptake values.

[0037] Another factor that plays an important role in the accurate quantification of SUV, is patient weight. Higher patient weight is associated with worse image quality and hence increased uncertainty in the measurement of SUV. Because of this, measurements such as SUVmax tend to show a strong correlation with patient weight. By normalizing the data across these different patient weights, the method ensures that the final imaging data is standardized, allowing for more accurate comparisons and analyses across different centers, protocols, and patient weights. This normalization is critical for improving the reliability of nuclear imaging in multicenter studies and clinical practice.

[0038] In a further embodiment, the nuclear imaging data from positron emission tomography (PET) or single-photon emission computed tomography (SPECT), which are commonly used in clinical practice and research. The method involves normalizing image intensity across the multiple sources to reduce biases, ensuring that the images are comparable in terms of intensity values. For nuclear images, these values may be quantified in standard uptake values (SUVs) or Counts (CNTS). Additionally, the method includes quantifying the reduction in biases achieved by the normalization process, providing a measure of its effectiveness and validating the approach. In an embodiment, the method includes a quantification of bias. This quantification provides a measure of the effectiveness of the normalization method and can be used to validate the approach. Latent representations of the nuclear imaging data, for instance obtained in the bottleneck of an encoder-decoder architecture, can be used to quantify if scans from different technologies or patient groups cluster together. Here, applying the same encoder on the normalized imaging data should result in less pronounced clustering, as measured by measured such as mutual information.

[0039] In an embodiment, the method works on minimizing bias in data that is specifically used in the context of multi-center clinical trials. Here, potential biases introduced by differences in scanner technologies, scanning protocols and patient weight can make the downstream processing and analyses more difficult. The proposed method can be used as a preprocessing step that facilitates downstream analyses. Machine learning models trained on this preprocessed data can greatly benefit from the more homogeneous data, resulting in a better model performance on downstream tasks and enhanced generalizability. Preferably, the model performance on data from independent centers that are preprocessed in the same way should increase the relevant metrics for the downstream task by 5%, more preferably by 10% and most preferably by 20%. These downstream tasks can include patient stratification for treatment response prediction, biomarker detection, patient triaging to automatically assess if a disease is present, and prediction of best follow-up treatment. The normalization step can be applied at the participating centers or at a central data hub, prior to any of these downstream analyses.

[0040] In an embodiment, the normalized nuclear imaging data can be used for various downstream analysis tasks, such as computer-aided diagnosis, computer-aided prognosis, disease classification, treatment response prediction, biomarker detection, image segmentation, radiomics feature extraction, longitudinal patient monitoring, drug development assessment, clinical trial analysis, and epidemiological studies. The use of normalized data in these tasks yields improved accuracy compared to using non-normalized data, as the reduced biases lead to more reliable and consistent results.

[0041] In an embodiment, the deep learning algorithm comprises a neural network with at least one convolutional layers, and the neural network is referred to as a

[0042] Convolutional Neural Network (CNN). In an embodiment, the CNN is trained by sampling 3D volumes of voxels from high- signal, metabolically relevant regions, which ensures that the network learns the most relevant features for normalization.

[0043] In an embodiment, the deep learning algorithm includes an encoder network that is obtained by training a combined encoder-decoder network on one or more tasks using nuclear imaging data. These tasks may include a classification task, an autoencoding task, a denoising task and / or a segmentation task. In a subsequent step, the decoder portion of the trained network is removed thereby generating a remaining encoder network. The remaining encoder portion may be fine-tuned on the normalization task. The encoder network generates encoded embeddings of the nuclear imaging data, capturing relevant features for normalization while preferably preserving task-specific information from the initial task such as the classification task. The encoder network and / or embeddings are optimised to reduce inter-center, intra-center, and patient weight biases in the encoded embeddings. In an embodiment, the normalized nuclear imaging data is reconstructed from the bias- reduced encoded embeddings.

[0044] In an embodiment, said deep learning algorithm comprises an encoder network that is obtained by: training a combined encoder-decoder network on a classification task using the nuclear imaging data; removing the decoder portion of the trained network; and fine tuning the remaining encoder portion on the normalization task wherein the encoder network generates encoded embeddings of the nuclear imaging data; wherein the encoded embeddings capture relevant features for normalization while preserving task-specific information from the initial classification task; wherein the fine-tuning process optimizes the encoder network to reduce inter-center, intra- center, and patient weight biases in the encoded embeddings; and wherein the normalized nuclear imaging data is reconstructed from the bias-reduced encoded embeddings.

[0045] Moreover, the deep learning algorithm is configured to adapt to new sources of nuclear imaging data without retraining, for instance by transfer learning or domain adaptation. This adaptability ensures that the method remains effective even as new imaging centers, scanner models, or protocols are introduced, providing a robust and scalable solution for normalizing nuclear imaging data. The normalized data generated by this method ensures consistent and reliable interpretation of imaging results, which is crucial for accurate diagnosis and treatment planning, making it a valuable tool for clinical practice and research. The method preferably employs a deep learning algorithm to process the received nuclear imaging data. More preferably, the deep learning algorithm comprises convolutional neural network (CNN) layers. The CNN is preferably trained by sampling 3D volumes of voxels from metabolically relevant regions, ensuring that the algorithm is adept at handling the intricacies of nuclear imaging data.

[0046] The multiple sources comprise at least two different imaging origins selected from the group consisting of different imaging centers; different scanner hardware models or manufacturers; different reconstruction algorithms; different acquisition protocols; different radiotracer dosing schemes; different imaging time points postradiotracer administration; and patient weight.

[0047] Normalizing nuclear imaging data reduces variability and improves diagnostic reliability across different sources, leading to more accurate and consistent medical diagnoses. By addressing inter-center, intra-center, and patient weight biases, the method ensures that the imaging data is more uniform and comparable, regardless of the originating source, which is particularly beneficial in multi-center clinical trials.

[0048] Preferably, the method includes quantifying the reduction in biases achieved by normalization. This step allows for an objective assessment of the process's effectiveness and provides valuable feedback for further optimization. The normalized nuclear imaging data may be used for various downstream tasks, such as computer-aided diagnosis, disease classification, treatment response prediction, image segmentation, and clinical trial analysis, yielding improved accuracy and enhancing the reliability and utility of nuclear imaging in clinical and research settings.

[0049] The method is designed to handle nuclear imaging data from both positron emission tomography (PET) and single-photon emission computed tomography (SPECT), enhancing its versatility and applicability across a broad spectrum of scenarios. PET is known for its high sensitivity and quantitative accuracy in detecting metabolic activity, while SPECT provides detailed functional information at a lower cost. By normalizing data from these modalities, the method ensures consistent and comparable results, addressing specific challenges such as partial volume effects in PET and collimator resolution in SPECT. Furthermore, the method may be configured to handle multi-modal imaging studies involving both PET and SPECT scans of the same patient. This capability is particularly useful in clinical research and multi-center trials, where different modalities gather complementary information, leading to more robust outcomes. Additionally, the method includes normalizing image intensity across multiple sources, crucial for ensuring that normalized data is consistent and comparable. The normalized nuclear imaging data is preferably used for multi-center clinical trials, enhancing the consistency and reliability of imaging results across different centers. This normalization leads to more robust and reliable study outcomes and improves the accuracy of downstream analysis tasks, including computer-aided diagnosis, disease classification, treatment response prediction, biomarker detection, image segmentation, radiomics feature extraction, longitudinal patient monitoring, drug development assessment, clinical trial analysis, and epidemiological studies. The utilization of normalized data in these tasks yields improved accuracy compared to non-normalized data, underscoring the significance of the normalization method in advancing nuclear imaging analysis.

[0050] In a preferred embodiment, the method further comprises quantifying the reduction in biases achieved by the normalization. This quantification step validates the effectiveness of the normalization method, ensuring reliable and reproducible nuclear imaging data analysis. The process may involve statistical analysis, machine learning metrics, or other techniques to measure bias reduction, providing a numerical or categorical output indicating the level of normalization achieved. Metrics such as mean squared error, signal-to-noise ratio, or visual inspection by experts complement quantitative assessments, ensuring the high quality of normalized data for downstream tasks.

[0051] The quantification step may be integrated into the deep learning algorithm used for normalization, allowing the algorithm to self-evaluate and adjust its parameters to minimize biases further. This adaptive approach enhances the robustness and generalizability of the method, making it applicable to various nuclear imaging data sources. In some cases, the quantification of bias reduction is performed at multiple stages of the normalization process, ensuring consistent bias reduction throughout the workflow, from data acquisition to final output.

[0052] In an embodiment, the method involves receiving nuclear imaging data from multiple sources, processing it using a deep learning algorithm, reducing intercenter, intra-center, and patient weight biases, and outputting normalized nuclear imaging data. This approach addresses challenges associated with integrating nuclear imaging data from diverse sources, such as different imaging centers, scanner hardware models, reconstruction algorithms, acquisition protocols, radiotracer dosing schemes, imaging time points, and patient weight. By reducing biases from these varying factors, the method ensures the comparability and consistency of nuclear imaging data, which is crucial for accurate medical diagnoses and research outcomes.

[0053] It will be clear to a skilled person that the embodiments as discussed above are also applicable to this computer-implemented method.

[0054] The method's ability to adapt to new sources of nuclear imaging data without requiring retraining is another significant advantage. This adaptability ensures that the method remains robust and effective even as new imaging technologies and protocols emerge. The normalized nuclear imaging data can be used in various downstream analysis tasks, including computer-aided diagnosis, prognosis, disease classification, treatment response prediction, biomarker detection, image segmentation, radiomics feature extraction, longitudinal patient monitoring, drug development assessment, clinical trial analysis, and epidemiological studies. The utilization of normalized data in these tasks yields improved accuracy compared to using non-normalized data.

[0055] The method includes quantifying the reduction in biases achieved by the normalization process. This quantification provides a measurable indication of the effectiveness of the normalization, which can be critical for validating the method's performance in clinical and research settings. The normalized nuclear imaging data is particularly beneficial for multi-center clinical trials, where consistency and comparability of imaging data across different centers are paramount.

[0056] The method's ability to normalize image intensity across multiple sources is another key feature. This normalization ensures that the intensity values in the nuclear imaging data are consistent, regardless of the source, which is essential for accurate interpretation and analysis. The normalization process also addresses patient weight biases, ensuring that the imaging data is not unduly influenced by individual patient characteristics, thus providing a more accurate representation of the underlying biological processes. Overall, the method described in this aspect offers a comprehensive solution for normalizing nuclear imaging data, addressing the challenges posed by diverse data sources, and enabling more accurate and reliable analyses in medical and research applications. The method's adaptability, effectiveness, and ability to improve the accuracy of downstream analysis tasks make it a valuable tool in the field of nuclear imaging.

[0057] In a preferred embodiment, the deep learning algorithm comprises convolutional neural network (CNN) layers. Preferably, the CNN layers are configured to process the nuclear imaging data in a manner that enhances the extraction of relevant features while minimizing noise and artifacts. More preferably, the CNN layers are designed to handle the three-dimensional nature of the nuclear imaging data, ensuring that spatial relationships within the data are preserved and accurately represented. This approach may lead to more precise and clinically useful normalized nuclear imaging data.

[0058] The CNN layers are preferably trained on a diverse dataset of nuclear imaging data from various sources, focusing on high-signal, metabolically relevant regions to learn clinically significant features, thus improving the diagnostic accuracy of the normalized data. The training process ensures that the CNN layers generalize well to new, unseen data, enhancing the robustness and reliability of the normalized nuclear imaging data.

[0059] Preferably, the CNN layers are designed to capture intricate patterns and features in the nuclear imaging data, which may include positron emission tomography (PET) or single-photon emission computed tomography (SPECT) data. The CNN layers preferably consist of multiple convolutional, pooling, and fully connected layers that work together to process and normalize the imaging data.

[0060] Preferably, the CNN layers are trained using a large dataset of nuclear imaging data from multiple sources, which may include different imaging centers, scanner hardware models, reconstruction algorithms, acquisition protocols, radiotracer dosing schemes, imaging time points post-radiotracer administration, and patient population characteristics. This diverse training data helps the CNN layers to generalize well across different imaging conditions and patient morphologies, thereby reducing inter-center, intra-center, and patient weight biases. The precision of the deep learning model in normalizing nuclear imaging data is preferably quantified by evaluating the reduction in biases across different sources. This may involve comparing the normalized data with ground truth data or using statistical measures or p-values to assess the consistency and accuracy of the normalized data. A second technique to evaluate the model performance may include the deployment of an algorithm, for instance a deep learning model, to perform a classification task. The algorithm performance may then be increased when the normalized test data set is used for evaluation, compared to the unnormalized test data set.

[0061] In a preferred embodiment, the CNN layers are part of a deep learning algorithm that includes an encoder network. The encoder network is trained using a combined encoder-decoder network on a classification task with nuclear imaging data. After training, the decoder portion is removed, and the encoder is fine-tuned for the normalization task, optimizing it to reduce inter-center, intra-center, and patient weight biases while preserving essential task-specific information from the initial classification task.

[0062] This fine-tuning process allows the encoder network to focus on relevant features, enabling faster and more efficient processing of nuclear imaging data. By concentrating on these features, the network reduces computational time and resource usage while maintaining the quality and accuracy of the normalized data. Additionally, the fine-tuning process is adaptable, allowing the encoder network to generalize to new sources of nuclear imaging data without requiring complete retraining, which is particularly beneficial in multi-center clinical trials.

[0063] The method may also include quantifying the reduction in biases achieved by the normalization process, providing a measurable indication of the method's effectiveness. This is advantageous in clinical and research settings where data consistency is crucial, making the method valuable for applications such as computer-aided diagnosis, disease classification, and treatment response prediction.

[0064] In an embodiment, the deep learning algorithm is configured to adapt to new sources of nuclear imaging data without retraining, utilizing a modular architecture that facilitates the incorporation of new imaging data sources. This modularity allows for the addition of data processing modules to handle specific characteristics of new sources, such as different scanner hardware models or novel radiotracer dosing schemes. The algorithm may also employ transfer learning techniques, leveraging pre-existing knowledge to efficiently adapt to new data sources and maintain high performance levels without extensive retraining.

[0065] Preferably, the algorithm includes continuous learning mechanisms that update its parameters incrementally as new data becomes available, ensuring it remains up- to-date with the latest imaging data. This continuous learning approach operates in the background, minimizing disruption to clinical workflows while further enhancing the algorithm's ability to reduce biases and maintain consistent performance across diverse datasets.

[0066] The ability to adapt to new sources of nuclear imaging data without retraining is particularly beneficial for multi-center clinical trials. In such trials, imaging data is often collected from diverse sources, and the consistency of the normalized data is crucial for accurate analysis. By maintaining consistent performance across different datasets, the method ensures reliable and reproducible results, thereby enhancing the validity of the clinical trial outcomes.

[0067] In a final aspect, the current invention also discloses a non-transitory computer- readable medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform the method as described herein. The computer-readable medium may comprise, for example, an optical disk, a magnetic disk, a flash memory device, or any other tangible storage medium suitable for storing program code. When executed, the program code implements the steps of receiving nuclear imaging data from one or more sources, applying a deep learning algorithm to mitigate inter-center, intra-center, and patient weight biases, and generating normalized nuclear imaging data with reduced biases. In some embodiments, the stored instructions further enable the system to quantify the reduction in bias, adapt to new imaging sources without retraining, and provide the normalized data for downstream applications such as computer-aided diagnosis, treatment response prediction, or multi-center clinical trials.

[0068] In another aspect, the present invention provides a system for normalizing nuclear imaging data, the system comprising one or more processors and a memory storing instructions that, when executed by the one or more processors, cause the system to perform any of the methods described herein. The system is configured to receive nuclear imaging scans from one or more patients or clinical phantoms, apply a deep learning algorithm to mitigate inter-center, intra-center, and patient weight biases, and generate normalized nuclear imaging data with reduced biases. In preferred embodiments, the deep learning algorithm comprises a convolutional neural network trained on 3D volumes of voxels from metabolically relevant regions, and may include an encoder network fine-tuned for normalization. The system may further be configured to quantify the reduction in biases, to adapt to new sources of nuclear imaging data without retraining, and to output the normalized nuclear imaging data for use in downstream tasks such as computer-aided diagnosis, biomarker detection, treatment response prediction, longitudinal patient monitoring, or multi-center clinical trials.

[0069] The invention will now be described by means of examples and figures which are not limitative for the invention.

[0070] FIGURES

[0071] Figure 1 illustrates a schematic overview of an embodiment of the invention in which heterogeneous nuclear imaging data originating from different imaging centers (Centers A-C) are provided as input to a harmonizer model. The harmonizer model applies deep learning-based normalization to mitigate inter-center, intracenter, and patient weight biases, thereby generating harmonized nuclear imaging data. The harmonized data is then supplied to a trained classifier, which performs a downstream analysis task, in this example, a lesion triage task to determine whether a tumor is present (yes / no). This figure demonstrates how the normalization step facilitates improved performance in subsequent classification tasks by reducing variability in the input data.

[0072] Figure 2 shows a scatter plot of standardized uptake value (SUVmax) as a function of weight for both original (grey circles) and harmonized (black crosses) nuclear imaging data. In the original data, SUVmax exhibits a strong dependency on weight, indicating the presence of a systematic bias. After processing with the harmonizer model, this dependency is substantially reduced, as shown by the flatter regression line for the harmonized data. This demonstrates that the normalization method effectively mitigates weight-related bias in SUV quantification, leading to more consistent and comparable measurements across heterogeneous input data. EXAMPLES

[0073] Example 1: Tumor triaging classification performance increased with deep learning PET image normalization

[0074] Aim / Introduction

[0075] Whole-body FDG PET / CT imaging is routinely employed worldwide across various clinical applications. It plays a crucial role in the diagnosis, staging, restaging, and monitoring of treatment response in numerous cancer types. Additionally, it is valuable in assessing infectious and inflammatory diseases. The interpretation of FDG PET / CT studies is a complex and time-consuming process, requiring careful differentiation between normal FDG biodistribution patterns— such as muscle, myocardial, and bowel uptake— and pathological patterns, including those associated with inflammation, healing, benign tumors, and other incidental findings. This complexity is influenced by the reader's experience and the intricacy of the study. The focus of this study was to evaluate whether deep learning-based normalization of heterogeneous input data— stemming from variations in scanning protocols and scanner settings— could enhance classification performance. Specifically, the study aimed to determine if such normalization could contribute to more consistent and accurate interpretations, ultimately supporting improved diagnostic accuracy and efficiency in whole-body FDG PET / CT imaging.

[0076] Materials and Methods

[0077] A total of 2100 whole-body FDG scans performed in the clinical setting in one of the 3 participating centers were included, equally distributed in 2 groups:

[0078] • Normal studies: scans classified as normal based on clinical reports, with no evidence of cancer or progressive systemic disease documented in the patient's medical file within the following six months.

[0079] • Abnormal studies: scans exhibiting any non-physiological FDG uptake, including benign or malignant tumors, inflammatory, infectious, or fibrotic processes, or unexplained FDG uptake that does not correspond to known normal variants.

[0080] The dataset was divided into 80% for training plus validation and 20% for testing. Initially, a deep learning model consisting of convolutional and fully connected layers was trained to perform a binary classification, categorizing scans as either normal or abnormal. Subsequently, the same model was applied to the test set after normalization by a second deep learning model, which adjusted for variations in scanning protocols and scanner settings, to evaluate performance differences postnormalization. This method is also schematized in Figure 1.

[0081] Results

[0082] Applying deep learning-based normalization prior to the classification task significantly enhanced the model's performance in triaging scans. As shown in Figure 3, the "Area under the ROC Curve" (AUC) increased from 0.85 without normalization to 0.91 with normalization, demonstrating a clear improvement in classification accuracy.

[0083] Conclusion

[0084] The study demonstrated that deep learning-based normalization of heterogeneous FDG PET / CT data significantly improves the accuracy of scan classification. By standardizing input data across different scanning protocols and settings, the normalization process enhanced the model's ability to distinguish between normal and abnormal studies. This suggests that such normalization can be a valuable tool in clinical practice, leading to more reliable diagnostic outcomes.

[0085] Example 2: mitigating heterogeneity in SUV quantification due to scanner technologies and acouisition protocols: a phantom study

[0086] Aim / Introduction

[0087] Over the past decade PET imaging has seen several technological improvements such as the advent of digital scanning, time-of-flight (TOF) modelling and improved image reconstruction by explicitly taking into account the point spread function (PSF). However, the drawback of these new technologies is that in clinical practice a wide variety of imaging technologies is now available. In literature, it has been reported that these differences in underlying scanning technology can result in biases in SUV quantification which in turn can impede correct diagnosis and prognosis. The differences in scanner technology often make it difficult or even impossible to compare findings from different centers. To reduce this bias, we propose a deep learning model that can normalize PET scans from different modalities.

[0088] Materials and Methods

[0089] First, a deep-learning model with a U-Net architecture was trained on matching degraded and standard-count whole-body PET scans from two different hospitals. In this way, the model learned how to infer a ground truth image from images that had been affected by noise due to altered scanning protocols. From three independent centers 18F-FDG PET / CT phantom scans were collected, where each center could use the reconstruction parameters used in their clinical practice. The obtained scans were normalized by the deep learning model and both SUVmax and SUVmean were analyzed for both the original and the enhanced scans.

[0090] Results

[0091] The deep learning model demonstrated significant efficacy in normalizing PET scans across different modalities. The model successfully reduced the variability in the recovery coefficients for both SUVmax and SUVmean values between scans from different centers. Specifically, after normalization, the inter-center variability in RC- SUVmax decreased by more than 50% (p < 0.01) and in C-SUVmean by 10% (p < 0.01). Figure 4 shows the RC-SUV max for the different centers and different sphere sizes of the phantom. The phantom studies revealed that the normalized images had SUV values that were statistically closer to the reference values obtained from standard-count scans, indicating a successful mitigation of the biases introduced by different reconstruction parameters. Additionally, qualitative assessment by experienced nuclear medicine physicians showed improved consistency in image quality and lesion detectability across the normalized scans.

[0092] Conclusion

[0093] The deep learning model effectively normalized PET scans from different modalities, reducing the biases in SUV quantification that are introduced by varying scanner technologies and reconstruction protocols. This normalization allows for more reliable comparisons of PET findings across different clinical settings, potentially leading to more consistent diagnoses and prognoses. The results suggest that implementing such a model in clinical practice could facilitate multi-center studies and improve the standardization of PET imaging, ultimately enhancing patient care. Further studies with larger datasets and clinical validation are recommended to confirm these findings and explore the model's application in routine clinical workflows.

Claims

CLAIMS1. A method for normalizing two or more nuclear imaging scans from one or more patients, comprising: receiving at least two nuclear imaging scans; applying a deep learning algorithm to process the scans and mitigate biases arising from variations in at least one of the following: acquisition protocols, reconstruction algorithms, reconstruction parameters, and patient weight; and generating normalized nuclear imaging data with reduced biases.

2. The method of claim 1, wherein the nuclear imaging data comprise at least two different imaging origins selected from the group consisting of: different imaging centers; different scanner hardware models or manufacturers; different reconstruction algorithms; different acquisition protocols; different radiotracer dosing schemes; different imaging time points post-radiotracer administration; and differences in patient weight.

3. The method of any of the previous claims, wherein the nuclear imaging data includes data from at least one of: positron emission tomography (PET), or single-photon emission computed tomography (SPECT).

4. The method of any of the previous claims, wherein reducing biases comprises normalizing image intensity across the multiple sources.

5. The method of any of the previous claims, wherein the method further comprises quantifying the reduction in biases achieved by the normalization.

6. The method of any of the previous claims, wherein the normalized nuclear imaging data is used for multi-center clinical trials.

7. The method of any of the previous claims, further comprising: utilizing the normalized nuclear imaging data in at least one downstream analysis task selected from the group consisting of: computer-aided diagnosis; computer- aided prognosis; disease classification; treatment response prediction; biomarker detection; image segmentation; radiomics feature extraction; longitudinal patient monitoring; drug development assessment; clinical trial analysis; and epidemiological studies; and wherein the utilization of the normalized nuclear imaging data yields improved accuracy in the selecteddownstream analysis task compared to using non-normalized nuclear imaging data.

8. The method of any of the previous claims, wherein the deep learning algorithm is a neural network that comprises one or more convolutional layers.

9. The method of claim 8, wherein the neural network was trained by sampling 3D volumes of voxels from high-signal, metabolically relevant regions.

10. The method of any of the previous claims where normalized nuclear imaging scans are reconstructed from the bias-reduced encoded embeddings.

11. The method of any of the previous claims, wherein the deep learning algorithm comprises an encoder network that is obtained by:- training a combined encoder-decoder network on one or more tasks using the nuclear imaging data, such as a classification task, an auto-encoding task, a denoising task and / or a segmentation task ;- removing the decoder portion of the trained network; wherein the encoder network generates encoded embeddings of the nuclear imaging data that capture relevant features for normalization while preserving task-specific information from the one or more initial tasks; and wherein the encoder network and / or the embeddings are optimized to reduce inter-center, intracenter, and patient weight biases in the encoded embeddings.

12. The method according to claim 11, wherein said method comprises a step of fine-tuning a remaining encoder portion on the normalization task such that the encoder network generates encoded embeddings of the nuclear imaging data.

13. The method according to claim 11 or 12, wherein the normalized nuclear imaging data is reconstructed from the bias-reduced encoded embeddings.

14. The method of any of the previous claims, wherein the deep learning algorithm is configured to adapt to new sources of nuclear imaging data without retraining.

15. A non-transitory computer-readable medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform the method of any one of claims 1-14.

Images