Systems, devices and methods for harmonization of imaging data sets including biomarkers - Patents.com

Abstract

An exemplary system, method, and computer-accessible medium for harmonizing neuromelanin (NM) data using Combat directly on a NM database or for harmonizing future data using coefficients generated in Combat can include, for example, receiving imaging information of the brain of a patient(s) from one MRI scanner, receiving imaging information of the brain of a patient(s) from a second MRI scanner, and using Combat to harmonize the data between the scanners against a reference data set. The neuromelanin (NM) concentration of the patient(s) can then be determined based on the harmonized data. The NM concentration can be determined using a voxel-wise analysis procedure. The voxel-wise analysis procedure can be used to determine the tissue distribution pattern(s) within the substantia nigra (SN) of the brain of the patient(s).

Description

[Technical field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims priority to U.S. Provisional Application No. 63 / 159,915, entitled "CROSS-SCANNER HARMONIZATION OF NEUROMELANIN-SENSITIVE MRI FOR MULTISITE STUDIES," filed March 11, 2021, the entire disclosure of which is incorporated herein by reference.

[0002] Statement on Federally Sponsored Research This invention was made with Government support awarded under Grant Nos. R01-MH114965 and R01-MH117323 by the National Institutes of Health. The U.S. Government has certain rights in this invention.

[0003] Embodiments described herein relate to methods and apparatus for generating and / or correcting data associated with magnetic resonance imaging, and in particular to providing harmonized data associated with multiple sources of magnetic resonance imaging based on one or more biomarkers to improve accuracy rates in analysis based on multiple sources of magnetic resonance imaging. [Background technology]

[0004] The development and use of biomarkers is a promising approach to support the use of magnetic resonance imaging techniques to explore biological systems. Neuromelanin, an example of a biomarker, can be used as a dye or indicator to identify one or more biomarkers of neurodegeneration, such as in the dopaminergic system, via neuromelanin-sensitive magnetic resonance imaging (NM-MRI). The development and use of biomarkers typically involves the use of large sample sizes of scan data obtained from several multi-site MRI studies performed using different hardware and / or software. Combining scan data from multiple sites can introduce undesirable variability in the data due to differences in hardware and / or software. Thus, there is a need for a method to combine imaging or scan data from multiple sites without introducing excessive undesirable variability. Summary of the Invention

[0005] In some embodiments, the systems, devices, and methods described herein can assist scientists, physicians, researchers, clinicians, and other medical or clinical professionals in collecting and using large imaging datasets across multiple sources to perform robust research and draw powerful inferences based on the scan data.

[0006] Disclosed embodiments include a method of harmonizing a dataset that includes measuring a first level of neuromelanin associated with a first magnetic resonance imaging (MRI) device. The method includes measuring a second level of neuromelanin associated with a second MRI device. The method includes generating a set of input feature vectors based on the neuromelanin dataset. The method includes providing the set of input feature vectors to a machine learning model for adjusting the first level of neuromelanin associated with the first MRI device based on the second level of neuromelanin associated with the second MRI device.

[0007] Disclosed embodiments include an apparatus having a memory and a processor operatively coupled to the memory. The processor is configured to measure a set of first levels of neuromelanin at a set of time points using a first MRI device. The processor is configured to measure a second level of neuromelanin using a second reference MRI device that is different from the first MRI device. The processor is configured to implement a machine learning (ComBat) algorithm to generate scanner-specific coefficients associated with the first MRI device and the second reference MRI device. The scanner-specific coefficients are configured to define a relationship between the first MRI device and the second reference MRI device. The processor is configured to harmonize a set of data collected using the first MRI device compared to the second reference MRI device using the scanner-specific coefficients, the set of data including several NM scans.

[0008] Disclosed embodiments include a method of harmonization of neuromelanin data sets using vendor-specific combat coefficients. The method includes measuring a first level of neuromelanin associated with a first data set acquired using a set of MRI devices manufactured by a first vendor and measuring a second level of neuromelanin associated with a second data set acquired using a reference MRI device manufactured by a second vendor different from the first vendor. The method includes implementing a machine learning (combat) algorithm to generate vendor-specific coefficients associated with the set of MRI devices, the vendor-specific coefficients being based on at least one of (1) the first level of neuromelanin associated with the set of MRI devices, or (2) the second level of neuromelanin associated with the reference MRI device. The method includes acquiring a third data set associated with the first MRI device manufactured by the first vendor. The method includes applying the vendor-specific coefficients directly to the third data set from the first MRI device manufactured by the first vendor to measure an adjusted third level of neuromelanin associated with the first MRI device compared to at least one of (1) the second level of neuromelanin associated with the reference MRI device, or (2) the second data set. [Brief description of the drawings]

[0009] [Figure 1] FIG. 1 is a schematic diagram of a data harmonization system (DH system) according to one embodiment. [Diagram 2] FIG. 1 is a schematic diagram of a computing device in a data harmonization system, according to one embodiment. [Diagram 3] FIG. 1 is a schematic diagram of an analysis device in a data harmonization system, according to one embodiment. [Figure 4] FIG. 1 is a schematic diagram depicting an exemplary implementation of a data harmonization system, according to one embodiment. [Diagram 5]1 is a flowchart illustrating a method for harmonizing scan data associated with two sources using a data harmonization system, according to one embodiment. [Figure 6] 1 is a flow chart illustrating a method for initiating pre-processing of information for imaging data harmonization using a data harmonization system, according to one embodiment. [Figure 7] 1 is a flow chart illustrating a method for pre-processing information related to harmonization of imaging data using a data harmonization system, according to one embodiment. [Figure 8] 1 is a flow chart illustrating a method for harmonization of imaging data using a data harmonization system, according to one embodiment. [Figure 9] 1A and 1B are graphical representations of exemplary imaging data obtained from neuromelanin-sensitive magnetic resonance imaging (NM-MRI) using multiple imaging sources, before and after harmonization, respectively, using a data harmonization system, according to one embodiment. [Figure 10] 1A and 1B are graphical representations of exemplary imaging data obtained from an NM-MRI using multiple imaging sources, before and after harmonization, respectively, using a data harmonization system, according to one embodiment. [Figure 11] 1 is a graphical representation of imaging source classification based on exemplary imaging data obtained from NM-MRI using multiple imaging sources, showing imaging data-based classification performance before and after harmonization using a data harmonization system, according to one embodiment. [Figure 12] 1A and 1B are graphical representations of biologically meaningful variability associated with exemplary imaging data obtained from NM-MRI using multiple imaging sources, before and after harmonization using a data harmonization system, respectively, in accordance with one embodiment. [Figure 13]1A and 1B are images representing exemplary imaging data obtained from an NM-MRI using multiple imaging sources, before and after harmonization, respectively, using a data harmonization system, according to one embodiment. [Figure 14] 1A-1C are graphical representations of biologically meaningful information in exemplary imaging data obtained from NM-MRI using multiple imaging sources, before and after harmonization using a data harmonization system, according to one embodiment. [Figure 15A] 13A-13C are graphical representations of the reproducibility of information associated with biological data in exemplary imaging data obtained from NM-MRI using multiple imaging sources, before and after harmonization using a data harmonization system, according to one embodiment. [Figure 15B] 1 is a graphical representation of the overlap of information associated with biological data in exemplary imaging data obtained from NM-MRI using multiple imaging sources, before and after harmonization using a data harmonization system, according to one embodiment. [Figure 16] 1A and 1B are graphical representations of exemplary imaging data obtained from NM-MRI using multiple imaging sources, before and after harmonization, respectively. [Figure 17] 1A and 1B are graphical representations of exemplary imaging data acquired using NM-MRI with two categories of imaging sources, respectively, after harmonization using a data harmonization system, according to one embodiment. The graphs show harmonization based on data sets of different sizes. [Figure 18] 1A and 1B are graphical representations of the harmonization measures achieved with different sized datasets obtained using NM-MRI with multiple imaging sources of the first and second categories, respectively, using a data harmonization system according to one embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] Magnetic resonance imaging provides a non-invasive approach to explore various biological systems for research and treatment. Image analysis can provide qualitative and / or quantitative information about the biological system based on one or more imaging biomarkers. Imaging biomarkers can be objective features or characteristics that are consistently observed in in vivo images and can be reproducibly used as indicators of the presence of a particular biological process, pathogenesis, or response to an intervention (e.g., therapeutic manipulation). In some cases, biomarkers can be used to derive a quantitative assessment of biological tissue. For example, biomarkers can be measured on a relative scale, such as a proportional or interval (difference) scale of two image values ​​(e.g., intensity) that can be meaningfully interpreted. In some cases, biomarkers can be measured on an interval scale. A wide range of properties of biological tissues can be investigated via MRI-based imaging by using biomarkers, including tissue microstructure, metabolism, composition, function, and morphology.

[0011] Despite their potential great utility and high demand, surprisingly few MRI biomarkers have been adopted for routine clinical use. The search for biomarkers in neuropsychiatric disorders is a major challenge in the field and dovetails with ongoing precision medicine efforts.

[0012] Biomarkers can be targeted by biomarker-specific imaging methods. Neuromelanin is an example of a biomarker or biomarker indication that can be widely used in research, diagnostic, clinical, and other medical endeavors via neuromelanin-sensitive magnetic resonance imaging (NM-MRI). NM-MRI has been shown to serve as a validated measure of neuromelanin concentration in the substantia nigra-ventral tegmental area (SN-VTA) complex and can be a surrogate measure of dopaminergic function with good test-retest reliability. Given the importance of the dopaminergic system in both basic and clinical neuroscience, NM-MRI has great potential as a non-invasive biomarker. Development of generalizable biomarkers typically requires large data samples collected across multiple studies and / or sites. However, as described, combining imaging data obtained from multiple sources, e.g., NM-MRI data, can introduce source-induced variability that can mislead studies.

[0013] The embodiments disclosed herein include devices and methods to overcome challenges faced in large-scale integration of data from multiple studies / sites through harmonization of MRI signals based on biomarkers acquired from multiple sources. The examples disclosed herein show successful harmonization of NM-MRI data obtained from 128 healthy subjects from three different sites and acquired using five different MRI scanners. Harmonization of the imaging data using a machine learning algorithm (ComBat) effectively reduced significant differences due to variability in SN-VTA NM-MRI signals between MRI scanners while maintaining biologically meaningful variability associated with age effects. Harmonization also enhanced reproducibility and statistical power of known biologically meaningful age effects. The devices and methods presented herein can be useful for large-scale multi-site studies to develop biomarkers based on NM-MRI.

[0014] NM-MRI has been validated as a marker of both dopaminergic function and dopaminergic neurodegeneration, demonstrating its potential as a biomarker of the dopaminergic system. Furthermore, NM-MRI has several features that position it as an ideal biomarker candidate, including its non-invasiveness and lack of ionizing radiation allowing repeated and longitudinal measurements, its ease of implementation, and its excellent test-retest reliability. Biomarker development and characterization typically involves the use of large sample sizes to facilitate training and testing of statistical models and provide generalizability. Although multi-site studies facilitate obtaining such large sample sizes, methods to combine NM-MRI data from multiple sites have yet to be developed.

[0015] Combining imaging data from multiple sites can introduce unwanted nonbiological variability into the data due to differences in hardware (e.g., MRI scanner or head coil) or software (e.g., pulse sequence parameters). Including covariates in the analysis (i.e., statistically controlling for site) may not differ from models that ignore confounding if it is done without adequately removing this variance.

[0016] Conventional harmonization methods for generalized data, i.e., conventional explicit removal of non-biological variability, include a machine learning algorithm called ComBat that has emerged as a standard in the literature. The ComBat algorithm is an empirical Bayesian method for harmonizing both the mean and variance across batches of data, and is particularly robust for handling datasets with small sample sizes. Approaches using the ComBat algorithm have previously been applied to MRI-based measures of cortical thickness, regional brain volume and cortical surface area, diffusion tensor imaging, resting-state functional MRI, and task-based functional MRI. However, there is no established method for applying the ComBat algorithm to biomarker-specific imaging datasets. For example, there is an unmet need for a method to apply the ComBat algorithm to the harmonization of neuromelanin-specific imaging data acquired via NM-MRI.

[0017] The systems, devices, and methods disclosed herein can implement a harmonization method for biomarker-specific imaging data from multiple MRI sources across multiple imaging sessions, where acquisition parameters were not optimized across imaging sessions. For example, the systems, devices, and methods disclosed herein can implement a harmonization method for NM-MRI data, where the NM-MRI data was acquired from multiple patients, imaged across multiple sites, imaged using multiple MRI machines, and / or imaged using multiple MRI machines from multiple vendors, such as GE, Siemens, and Philips, where acquisition parameters across the various imaging sessions were not optimized. The systems, devices, and methods disclosed herein can implement a method for harmonizing the NM-MRI data while preserving biologically meaningful variability, such as variability associated with age. The examples presented herein show successful harmonization of NM-MRI data acquired from three major MRI vendors (GE, Siemens, and Philips) while preserving biologically meaningful variability associated with the increase in NM-MRI signal with age. The examples also show that harmonization improved both the reproducibility and statistical power of the predicted positive relationship between NM-MRI signal and age.

[0018] FIG. 1 is a schematic diagram of a system 100, which may be a data harmonization system (also referred to herein as a "DH system" or "system"). The system 100 includes a set of computing devices, including, for example, an analysis device 105 (e.g., an analysis device configured for data harmonization), a database 104 (e.g., a database that includes or stores specific MRI images associated with one or more biomarkers, one or more patients, MRI machines from one or more vendors, etc.), and one or more other computing device(s) 101, 103. The system 100 can assist in the harmonization of biomarker-specific imaging data provided by a user (e.g., a physician, radiologist, diagnostician, clinician, research professional, etc.) and / or a computer-aided diagnostic device. The analysis device 105, the physician device 103, and / or the computing device(s) 101 can communicate with each other via a communication network 106, as shown in FIG. 1.

[0019] The analysis device 105 can receive biomarker-specific imaging data associated with one or more sources from any suitable device (e.g., computing device 101 or physician device 103) or database (e.g., database 104), and requests to harmonize the data with a standard reference or with each other if data from multiple sources is included. The imaging data can include images acquired via biomarker-specific MRI (e.g., NM-MRI) as well as information associated with the source (e.g., identity of the device sending the harmonization request, identity of the imaging source associated with the imaging data including the patient, imaging time of each data set, imaging site associated with each data set, equipment used, acquisition parameters used during each imaging session, etc.). The analysis device 105 can implement algorithms (e.g., ComBat algorithms) to pre-process the imaging data and / or harmonize the imaging data upon request. The harmonized imaging data can be provided to the device, e.g., computing device(s) 101 and / or physician device 103, in the same format or in an appropriately modified format, according to one embodiment.

[0020] In some embodiments, the communication network 106 (also referred to as a "network") can be any suitable communication network for transferring data, operating over a public network and / or a private network. For example, the network 106 can include a private network, a virtual private network (VPN), a multiprotocol label switching (MPLS) circuit, the Internet, an intranet, a local area network (LAN), a wide area network (WAN), a metropolitan area network (MAN), a Worldwide Interoperability of Microwave Access Networks (WiMAX), a fiber optic (or optical fiber) based network, a Bluetooth network, a virtual network, and / or any combination thereof. In some cases, the communication network 106 can be a wireless network, such as, for example, a Wi-Fi or wireless local area network ("WLAN"), a wireless wide area network ("WWAN"), and / or a cellular network. In other examples, the communication network 106 can be a wired network, such as, for example, an Ethernet network, a digital subscriber line ("DSL") network, a broadband network, and / or an optical fiber network. In some cases, the network may use application programming interfaces (APIs) and / or data exchange formats (e.g., Representational State Transfer (REST), JavaScript Object Notation (JSON), Extensible Markup Language (XML), Simple Object Access Protocol (SOAP), and / or Java Message Service (JMS)). Communications transmitted over the network 106 may be encrypted or decrypted.In some cases, communications network 106 may include multiple networks or sub-networks operatively coupled to one another, for example, by network bridges, routers, switches, gateways, etc. (not shown).

[0021] The computing devices in the system 100, including, for example, the analysis device 105, the physician device 103, and / or the other computing device(s) 101, can each be any suitable hardware-based computing device and / or multimedia device, such as, for example, a server, a desktop computing device, a smartphone, a tablet, a wearable device, a laptop, etc. The physician device 103 can be, for example, a workstation or other device, such as an MRI scanner, used by a physician, radiologist, diagnostician, researcher, or other individual associated with the imaging data set. For example, in some embodiments, the physician device 103 can include a radiology workstation associated with a radiologist and / or a radiology service providing entity. In some cases, the radiology workstation can be equipped to communicate with the system 100 to obtain radiology imaging data of samples from a patient and / or a medical institution and transmit the scanned imaging data to the analysis device 105, the database 104, and / or the other computing device(s) 101. In some cases, the imaging data can be appropriately updated (e.g., annotated) and / or transmitted from the radiology workstation via a DICOM viewer application. In some cases, the imaging data can be appropriately updated (e.g., annotated) and / or transmitted from the radiology workstation via a DICOM viewer application. The other computing device(s) 101 can include other devices that are part of the data harmonization system, including, for example, computing systems (e.g., systems including devices that implement computational algorithms, systems that provide computational services such as cloud-based computing frameworks, etc.), workstations, servers, user or patient devices, etc.

[0022] The system 100 may also include a database 104 for storing data related to, for example, patients, hospitals, imaging systems, etc. In some cases, the database 104 may include one or more devices that may be part of a picture archiving and communication system (PACS), such as a PACS server. Some devices of a PACS system, such as a PACS server, may be configured for digital storage, transmission, and retrieval of medical images, such as radiology images. A PACS server may include software and / or hardware components that directly interface with imaging modalities. Images may be transferred from the PACS server to one or more computing device(s) 101 (e.g., CAD devices) for display and / or reporting, and / or to an analysis device 105 for analysis. In some cases, the computing device 101 (e.g., CAD device) may access an image from the database 104 (e.g., PACS server) and send it to the analysis device 105. In some cases, the computing device 101 (e.g., CAD device) may retrieve an image directly from a physician device 103 (e.g., a radiology workstation) and send it to the analysis device 105.

[0023] 2 is a schematic block diagram of an exemplary computing device 201 that can be part of a data harmonization system (e.g., system 100), according to one embodiment. The computing device 201 can be structurally and / or functionally similar to the computing device(s) 101 and / or physician device 103 of the system 100 shown in FIG. 1. The computing device 201 can be a hardware-based computing device and / or a multimedia device, such as, for example, a server, a desktop computing device, a smartphone, a tablet, a wearable device, a laptop, an ultrasound scanner, etc. The computing device 201 includes a processor 211, a memory 212 (e.g., including data storage), and a communicator 213. In some embodiments, the computing device 201 can have any suitable number of additional components (not shown), including input / output devices, displays, etc.

[0024] The processor 211 can be, for example, a hardware-based integrated circuit (IC) or any other suitable processing device configured to perform and / or execute a set of instructions or code. For example, the processor 211 can be a general-purpose processor, a central processing unit (CPU), an accelerated processing unit (APU), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a programmable logic array (PLA), a complex programmable logic device (CPLD), a programmable logic controller (PLC), etc. The processor 211 can be operatively coupled to the memory 212 via a system bus (e.g., an address bus, a data bus, and / or a control bus). As described herein, the computing device 201 can be associated with and / or operatively coupled to an imaging device (e.g., an MRI device) and configured to acquire imaging data acquired by the device over one or more imaging sessions based on specified image acquisition parameters set at the imaging device.

[0025] The processor 211 may be configured to receive and / or obtain imaging data, metadata, clinical information, medical information, other information, etc., from a remote source (e.g., a clinical database (e.g., database 104), a repository, a scanner, an imaging device, other devices associated with radiology services, devices associated with a patient management system / hospital, devices associated with an individual / patient, etc.). In some cases, the processor 211 may be configured to receive imaging data and / or data related to the imaging data (e.g., processed data, measurements related to the imaging data, annotations related to the imaging data, historical / clinical data related to the imaging data, etc.) from any suitable source including an imaging (e.g., an MRI machine), a workstation, a physician's device, a patient device, a database, or any other suitable device.

[0026] The processor 211 can receive commands to process the data (e.g., can receive commands to harmonize the data). In some cases, the processor 211 can implement a user interface on the computing device 201 to receive commands, e.g., from a user, as described in further detail herein. For example, the processor 211 can be configured to cause an input / output device (e.g., an input / output device coupled to the communicator 213 or integrated into the computing device 201) to display an interface to an operator (e.g., a physician).

[0027] The processor 211 can respond to one or more commands to pre-process, organize, and / or initiate harmonization of the data. In some cases, the processor 211 of the computing device 201 can be configured to automatically identify a source of the data or determine the lack of an identified source of the data.

[0028] In some cases, the processor 211 can pre-process the data, send it to an analysis device, and initiate harmonization of the data, as described herein. In some cases, the processor 211 can be configured to receive the harmonized data, and in some embodiments, check the harmonization status of the data set. The processor 211 can be configured to present information related to the imaging data set to a user (e.g., information indicative of a measure of harmonization, or information indicative of retention of biological variability, etc.).

[0029] The processor 211 may include or execute one or more module(s) or instruction(s) stored in the memory 212 to function as the data handler 214 and the interface manager 215. In some embodiments, the processor 211 may include a software application (not shown in FIG. 2) that may be used to provide an interface for delivering a diagnostic assessment. The software application may be any suitable software or code that, when executed by the processor 211, may be configured to perform a coordinated set of functions, tasks, or activities for the benefit of a user of the computing device 201. The software application may be, for example, a browser application, a word processing application, a media playback application, a JAVA-based application, an image rendering or editing application, a text editing application, etc.

[0030] In some embodiments, each of the data handler 214, and interface manager 215, and / or software applications can be software stored in memory 212 and executed by processor 211. For example, each of the above-mentioned portions of processor 211 can be implemented in the form of instructions, which can include code configured to cause processor 211 to execute the data handler 214, and interface manager 215, and / or software applications. The code can be stored in memory 212 and / or in hardware-based devices, such as, for example, ASICs, FPGAs, CPLDs, PLAs, PLCs, etc. In other embodiments, each of the data handler 214, and interface manager 215, and / or software applications can be hardware configured to perform their respective functions.

[0031] The processor 211 implementing the data handler 214 can be configured to collect imaging data, annotation data, clinical data, and / or any other relevant data related to the patient, the site, the imaging session, the scanning equipment, the medical / clinical institution, and / or the imaging study being performed, etc. For example, the data handler 214 can receive, collect, process, and / or store imaging data, such as radiology data provided for the investigation, diagnosis, or evaluation of a biological system of interest.

[0032] The imaging data may be in the form of multiple images. In some cases, the multiple images can be acquired in a spatial sequence (e.g., successive spatial locations of the biological tissue according to a spatial order). In some cases, the multiple images can be acquired in a time sequence (e.g., successive time windows imaging the biological tissue according to a time order). In some cases, the multiple images can be acquired in a spatio-temporal sequence having a spatio-temporal order. The imaging data can also include metadata or supplemental details in any suitable format associated with the imaging data that provide additional information related to the circumstances under which the imaging data was acquired, such as the addition of the imaging device and / or the patient being imaged. For example, the imaging data can include metadata or supplemental details in the form of information related to image acquisition parameters, information related to the equipment used (such as imaging device vendor details such as make, model, date and batch of manufacture), information related to the calibration of the device, and the like. In some cases, the imaging data can be in the form of images obtained from a biomarker-specific MRI (e.g., NM-MRI). In some cases, the imaging data can be in the form of DICOM images and associated metadata.

[0033] The data handler 214 may also obtain clinical data, annotation-based data, or other supplemental data associated with the imaging data, one or more patient(s), image site, imaging window (i.e., time / day of imaging), biological condition, etc., to organize the data. In some cases, the clinical data may include, for example, histopathological data associated with the imaged biological tissue, medical information, biographical information associated with the subject or patient, information regarding other details of the patient having age, sex, ethnicity, health condition, medical history, etc. In some cases, the data handler 214 may determine the absence of specified information (e.g., the absence of the identity of the scanner associated with the data set) and await user input via an interface, for example, implemented by the interface manager 215, to provide information regarding the identity of the scanner associated with the data set. In some implementations, the data handler 214 and / or the interface manager 215 may also prompt the user to provide one or more specified inputs to record information associated with the imaging data set. In some implementations, the imaging data may be suitably encrypted, for example, to protect personal data associated with patients, subjects, social organizations, clinical trials, etc. In some implementations, the data handler 214 may be configured to decrypt the encrypted data as needed to extract information related to the imaging data set. The data handler 214 may also be configured to appropriately encrypt the information before transmitting it to another device (eg, an analysis device, a database, a computing device, etc.).

[0034] In some implementations, the data handler 214 can be configured to evaluate the identified imaging data to determine whether it has been harmonized. In some cases, the data handler 214 can be configured to receive a query, e.g., from a user via an interface managed by the interface manager 215, and in response to the query, evaluate the identified imaging data set(s) to whether it has been harmonized. In some implementations, the data handler 214 can be configured to automatically evaluate the identified imaging data to determine whether it has been harmonized, e.g., by comparing the identified imaging data set(s) to a predefined reference imaging data set(s).

[0035] In some implementations, the processor 211 implementing the data handler 214 can be configured to receive instructions from an operator or physician via an interface (e.g., a user interface implemented and managed by the interface manager 215, etc.) to harmonize one or more data sets according to predetermined conditions. The data handler 214 can organize the data necessary to harmonize one or more data sets according to predetermined conditions and provide the data to an analysis device (e.g., the analysis device 105 of FIG. 1). In some embodiments, the processor 211 can implement the data handler 214 as a background process that monitors for newly incoming data, processes the data, and / or stores the data.

[0036] The interface manager 215 can be configured to provide an interface (also referred to herein as a “user interface”) to an operator of the computing device 201 through which the operator may load or access the imaging data or any associated supplemental data, evaluate the imaging data or supplemental data, and / or initiate data harmonization. In some implementations, the interface manager 215 can provide the user with control options through the interface, such as to, for example, assign or organize a data structure to the imaging data and / or supplemental data, change and / or customize settings in the interface to organize the imaging data and / or supplemental data, etc. In some implementations, the interface manager 215 can be configured to provide the operator or user with control tools to add supplemental information that may not already be associated with the imaging data. For example, the interface manager 215 can enable the user to enter information related to the source of the imaging data set (e.g., biographical or patient background information about the patient, information about the imaging time, imaging site, imaging equipment, etc.). In some cases, the interface manager 215 can be configured to provide the user with tools to initiate harmonization of the imaging data set, as described herein. For example, the interface manager 215 can receive instructions from a user to initiate harmonization of the imaging dataset and, in response to those instructions, send a signal to an analysis device (e.g., an analysis device 105, 205 described herein) to perform the harmonization.

[0037] Interface manager 215 can be configured to receive harmonized imaging data, for example, from an analysis device (e.g., analysis device 105, 205) and provide the harmonized imaging data for further processing, review, analysis, evaluation, storage, etc.

[0038] Although described herein as being performed by a computing device 201 that can be separate from the analysis device that performs the harmonization of the data set, in some embodiments, the computing device 201 implementing the data handler 214 and the interface manager 215 can be the same analysis device that performs the harmonization of the data set. In other words, one or more of the above processes described with respect to the processor 211, e.g., implementing the data handler 214 and / or the interface manager 215, can be performed by an analysis device (e.g., the analysis device 105 included in the system 100 shown in FIG. 1 and / or the analysis device 305 shown in FIG. 3 and described in more detail in the following section).

[0039] Although computing device 201 is described as implementing each of data handler 214 and interface manager 215 via processor 211, it can be understood that a computing device can be comprised of several instances of the above-mentioned units, components, and / or modules. Furthermore, the terms data handler and interface manager are provided for illustrative purposes to describe processes implemented, for example, by processor 211. Thus, one or more of these modules can be combined into a single module or generally referred to as a processor configured to perform that one or more processes or steps.

[0040] The memory 212 of the computing device 201 can be, for example, a random access memory (RAM), a memory buffer, a hard drive, a read only memory (ROM), an erasable programmable read only memory (EPROM), etc. The memory 212 can store one or more software modules and / or code, which can include, for example, instructions for the processor 211 (e.g., the data handler 214, the interface manager 215, and / or the software applications described above) to execute one or more processes, functions, etc. In some embodiments, the memory 212 can include an expandable storage unit that can be added to and used incrementally. In some implementations, the memory 212 can be a portable memory (e.g., a flash drive, a portable hard disk, etc.) that can be operably coupled to the processor 211. In other examples, the memory can be operably coupled to the computing device remotely. For example, a remote database server can function as a memory and be operably coupled to the computing device.

[0041] The communicator 213 can be a hardware device operatively coupled to the processor 211 and the memory 212, and / or software stored in the memory 212 executed by the processor 211. The communicator 213 can be, for example, a network interface card (NIC), a Wi-Fi™ module, a Bluetooth® module, and / or any other suitable wired and / or wireless communication device. Furthermore, the communicator 213 can include a switch, a router, a hub, and / or any other network device. The communicator 213 can be configured to connect the computing device 201 to a communication network (such as the communication network 106 shown in FIG. 1). In some cases, communicator 213 may be configured to connect to a communications network such as, for example, the Internet, an intranet, a local area network (LAN), a wide area network (WAN), a metropolitan area network (MAN), a Worldwide Interoperability of Microwave Access Networks (WiMAX®), a fiber optic (or optical fiber) based network, a Bluetooth® network, a virtual network, and / or any combination thereof.

[0042] In some cases, the communicator 213 can facilitate receiving and / or transmitting signals (e.g., a signal to initiate harmonization of a data set), data (e.g., imaging data, supplemental data, clinical data, etc.), files, and / or file sets over a communications network (e.g., communications network 106 of system 100 of FIG. 1). In some cases, the received files can be processed by processor 211 and / or stored in memory 212, as described in further detail herein. In some cases, as described above, the communicator 213 can be configured to transmit data collected and / or analyzed by the data handler 214 to an analysis device (e.g., analysis device 105) of a data harmonization system (e.g., system 100) to which the computing device 201 is connected. The communicator 213 can also be configured to transmit data collected and analyzed by the data handler 214, and any analysis results generated, to an analysis device of a data harmonization system to which the computing device 201 is connected.

[0043] Returning to Figure 1, computing device 101 connected to system 100 can be configured to communicate with analysis device 105 via communication network 106. Figure 3 is a schematic diagram of an analysis device 305 that is part of a data harmonization system (e.g., system 100), according to some embodiments. Analysis device 305 can be structurally and / or functionally similar to analysis device 105 of system 100 shown in Figure 1. Analysis device 305 includes a communicator 353, a memory 352, and a processor 351.

[0044] Similar to the communicator 213 in the computing device 201 of FIG. 2, the communicator 353 can be a hardware device operatively coupled to the processor 351 and the memory 352, and / or software stored in the memory 352 for execution by the processor 351. The communicator 353 can be, for example, a network interface card (NIC), a Wi-Fi™ module, a Bluetooth® module, and / or any other suitable wired and / or wireless communication device. Additionally, the communicator 353 can include a switch, a router, a hub, and / or any other suitable network device. The communicator 353 can be configured to connect the analysis device 305 to a communication network (such as the communication network 106 shown in FIG. 1). In some cases, communicator 353 may be configured to connect to a communications network such as, for example, the Internet, an intranet, a local area network (LAN), a wide area network (WAN), a metropolitan area network (MAN), a Worldwide Interoperability of Microwave Access Networks (WiMAX®), a fiber optic (or optical fiber) based network, a Bluetooth® network, a virtual network, and / or any combination thereof.

[0045] Memory 352 can be a random access memory (RAM), a memory buffer, a hard drive, a read only memory (ROM), an erasable programmable read only memory (EPROM), etc. Memory 352 can store data, such as imaging data and / or any supplemental data related to the imaging data. Memory 352 can store one or more software modules and / or code, which can include, for example, instructions for processor 351 to perform one or more processes, functions, etc. In some implementations, memory 352 can be a portable memory (e.g., a flash drive, a portable hard disk, etc.) that can be operably coupled to processor 351. In other examples, memory 352 can be operably coupled remotely to analysis device 305. For example, memory can be a remote database server operably coupled to analysis device 305 and its components and / or modules.

[0046] The processor 351 can be a hardware-based integrated circuit (IC) or any other suitable processing device configured to perform and / or execute a set of instructions or code. For example, the processor 351 can be a general-purpose processor, a central processing unit (CPU), a graphics processing unit (GPU), an accelerated processing unit (APU), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), a programmable logic array (PLA), a complex programmable logic device (CPLD), a programmable logic controller (PLC), etc. The processor 351 is operatively coupled to the memory 352 via a system bus (e.g., an address bus, a data bus, and / or a control bus). The processor 351 is operatively coupled to the communicator 353 via an appropriate connection or device, as will be described in more detail.

[0047] The processor 351 may include and / or be configured to execute several components, units and / or modules that may be configured to perform several functions, as described in more detail herein. The components may be hardware-based components (e.g., an integrated circuit (IC) or any other suitable processing device configured to perform and / or execute a set of instructions or code), or software-based components (executed by the server processor 352), or a combination of the two.

[0048] As shown in FIG. 3, the processor 351 can include or execute one or more module(s) or instruction(s) stored in the memory 352 to function as a data manager 354 and a harmonizer 358.

[0049] A data manager 354 in the processor 351 can be configured to receive communications between the analysis device 305 and a computing device connected to the analysis device 305 via a suitable communications network (e.g., computing devices 101-103 connected to the analysis device 105 via communications network 106 in the system 100 of FIG. 1). The data manager 354 is configured to receive information about the imaging data (e.g., biomarker-specific MRI data such as the NM-MRI dataset), supplemental data such as clinical data, historical data, annotations, etc. from the computing device, database, and / or remote source. In some implementations, the data manager 354 can receive data from the computing device by signals transmitted through an interface (e.g., a user interface implemented by an interface manager such as interface manager 215).

[0050] The data manager 354 can be configured to receive radiological imaging data from physicians, researchers, patients, and any other suitable users, the data being associated with an organ and / or organ system under clinical or therapeutic analysis or investigation. In some cases, the data manager 354 can receive information related to a patient's medical history, biographical information (e.g., age, sex, race, ethnicity, health history, behavioral habits, etc.). The data manager 354 can receive information related to a reference data set that can be used to harmonize any given imaging data set by reducing variability in the given data set. The imaging data and / or reference data can be in the form of DICOM images and associated metadata, or other imaging data. In some implementations, the imaging data and / or reference data can be suitably encrypted to protect personal data associated with, for example, a patient, subject, social organization, clinical trial, etc. In some implementations, the data manager 354 can be configured to decrypt the encrypted data as necessary to extract information related to the imaging data set. The data manager 354 can also be configured to suitably encrypt information before transmitting the information to another device (e.g., a portal device, a database, a computing device, etc.).

[0051] In some examples, the processor can include one or more machine learning models (not shown in FIG. 3 ) configured to receive information related to the identified imaging data, for example, in the form of feature vectors extracted by the data manager 354. The models can include, but are not limited to, deep neural networks, multi-layer perceptrons, random forests, support vector machines, and the like. The machine learning models can be configured to process the information and provide an output, for example, in the form of coefficients configured to reduce variability. In some cases, the processor can be configured to harmonize data obtained from a first source (e.g., imaging data obtained from a first subject and / or at a first time point using a first instrument) with data obtained from a second source (e.g., imaging data obtained from a second subject and / or at a second time point using a second instrument). The coefficients can be configured to define a relationship between the first source and the second source. In some cases, the second source can be a reference source used to harmonize multiple data sets. In some embodiments, the machine learning models can be configured to provide an output including one or more coefficients, functions (e.g., transformation functions), and the like for harmonizing data from different sources.

[0052] The machine learning model can be configured to process information to provide an output indicative of one or more classes associated with the imaging data. For example, the machine learning model can function as a class classifier and process information related to the imaging data to provide as an output a classification of non-biological sources of variation in the imaging data (e.g., a class of imaging equipment defined by the manufacturer, a model used to generate the data, etc.). In some implementations, the model can be used to provide an output indicative of a classification of biologically or historically meaningful sources of variation in the imaging data (e.g., a classification based on age, sex, race, etc.). In some implementations, the machine learning model can be used to evaluate before and / or after normalization or harmonization of the imaging dataset.

[0053] The processor 351 includes a harmonizer 356 configured to implement a routine or algorithm for harmonizing the identified imaging dataset by reducing or decreasing non-biological or non-historically meaningful sources of variability associated with the identified dataset (e.g., the ComBat algorithm configured to harmonize NM-MRI imaging datasets). Harmonization can include identifying one or more sources of non-biological variability (e.g., variability associated with the imaging equipment) and defining one or more coefficients that, when applied to the identified imaging dataset, can be configured to reduce the amount of non-biological variability associated with the dataset. The harmonizer 358 can be configured to receive the identified dataset and any supplemental data associated with the identified imaging dataset. The harmonizer 358 can evaluate the supplemental data to determine sources of undesirable variability associated with the identified imaging dataset. In some implementations, the sources of undesirable variability can be sources of non-biological variability. For example, the supplemental data can indicate sources associated with the imaging dataset. The sources can be an identified imaging device (e.g., an identified scanner or MRI machine), an identified make or model of the imaging or scanning device (e.g., MR machine manufacturer, such as GE, Phillips, Siemens, etc.), an identified patient, etc. In some embodiments, biological sources of variability (e.g., biomarker-specific MRI variability associated with age, sex, race, etc.) can be the focus of the study. In studies such as these, reducing non-biological sources of variability (e.g., originating from variation in the MRI scanner, image acquisition parameters, or scanner manufacturer, etc.) can be useful to accurately identify variability that is due to biological differences associated with samples, patients, and / or imaging datasets.The harmonizer 358 can be configured to execute a machine learning algorithm (e.g., the ComBat algorithm) to define one or more coefficients (e.g., ComBat coefficients) that, when applied to adjust or transform the imaging dataset, can generate a normalized or harmonized dataset that reduces variability from undesirable non-biological sources while maintaining variability associated with desired biological sources.

[0054] In some implementations, a source of undesired variability may be of biological origin, but may not be the source under investigation in a particular instance. For example, one source of variability that may be of biological origin may be the focus of the investigation (e.g., age), and a second source of variability that may be of biological origin (e.g., gender) may be undesirable in the given investigation and may need to be controlled. The harmonizer 358 may be configured to define coefficients configured to reduce variability from undesired sources. The harmonizer 358 may be configured to define coefficients configured to generate a harmonized data set that, when applied to the identified imaging data set, reduces variability associated with undesired biological sources while retaining variability associated with desirable biological sources.

[0055] The harmonizer 358 can be configured to receive a reference imaging dataset and generate coefficients for the identified imaging dataset based on the reference imaging dataset, so that when applying the coefficients, the identified imaging data can be harmonized with respect to the reference imaging dataset. The reference imaging dataset can be selected based on the variability that is reduced by the harmonization procedure.

[0056] In some implementations, the harmonizer 358 can generate the coefficients using raw imaging data. In some implementations, the harmonizer 358 can generate the coefficients using processed data based on the imaging data. For example, the harmonizer 358 can use pre-processed data that has been averaged, registered, re-aligned, and / or transformed in any suitable format, e.g., consistent with a reference image. In some implementations, the harmonizer 358 can use one or more masks over predefined regions of the imaging data to obtain one or more measures related to the biomarkers. The harmonizer 358 can then run an algorithm (e.g., the ComBat algorithm) over the measures to generate the coefficients. The measures can be calculated using any suitable process.

[0057] In some exemplary implementations of the systems and methods described herein, MRI data sets acquired using different scanners across various clinical studies can be standardized, allowing for meaningful one-to-one comparisons between data sets. In some embodiments, the harmonizer 358 can be configured to standardize biomarker measurements in different subjects that are measured with different MRI machines or with varying image acquisition parameters. For example, the harmonizer 358 can be configured to standardize NM measurements in different subjects measured with different MRI machines or with the same MRI machine with different parameters. In some embodiments, the measurements can be collected in a large-scale clinical trial setting. Data sets from different MRI machines are not initially comparable because parameters between the MRI machines are not optimized or designed to be consistent with each other. After all patients have completed the clinical trial, the data harmonization systems and methods described herein can be implemented, for example, by running the ComBat algorithm across all biomarker-specific measures (e.g., CNR-NM voxels in each patient's masked region). The implementation of the COMBAT algorithm adjusts for differences in noise generated by MRI scanner variability between each CNR-NM voxel and normalizes each CNR-NM voxel measurement to a reference standard MRI machine.

[0058] The harmonizer 358 can execute the ComBat algorithm on a database of measures derived from the voxels of the imaging data, as described herein. First, a subject can be subjected to NM-susceptible MRI, resulting in raw brain MRI images. After pre-processing, a mask is applied to the identified brain region, e.g., the substantia nigra-ventral tegmental area (SN-VTA) region of the image. A measure, e.g., contrast-to-noise ratio (CNR), is calculated between voxels in the mask of the SN-VTA region and voxels in a control region of the imaging data. The processed voxels are referred to as CNR-NM voxels. These CNR-NM voxels from each patient participating in the clinical trial are stored in a database. This database may include data from multiple patients who have undergone NM-MRI scans on different MRI machines.

[0059] In some exemplary embodiments, the input to the ComBat algorithm can be CNR-NM voxels, and the output generated by the ComBat algorithm can be referred to as ComBat harmonized contrast-to-noise ratio neuromelanin voxels, or CH-CNR-NM voxels. The CH-CNR-NM voxels are then stored in a database for each individual patient. This embodiment allows for standardization of a large database containing CNR-NM voxels measured on different MRI machines, or the same MRI machine with different parameters, by running the ComBat algorithm once at the end of a clinical trial. In such exemplary embodiments, NM voxels that have not been processed into CNR-NM voxels may not be appropriate inputs to ComBat, and running the ComBat algorithm on NM voxels alone may not produce the desired output.

[0060] For example, NM-MRI images preprocessed in the MNI template (established template) space can be used to estimate NM-MRI contrast ratio (CNR) maps. The NM-MRI CNR at each voxel is a function of the NM-MRI signal intensity (I V) in the cerebral peduncle (an area of ​​white matter tracts known to have the lowest NM content) (I CC ) can be calculated as the percentage of signal difference from

[0061]

number

[0062] CNR values ​​of voxels within the substantia nigra-ventral tegmental area (SN-VTA) complex mask can be harmonized using the ComBat harmonization model to remove non-biological variability (associated with MRI scanning equipment variations) while preserving biological variability (e.g., age- and sex-related variability). Other examples can include harmonizing outcome measures (e.g., SN-VTA CNR or regional brain volume) instead of raw MRI signals. The model can be written as follows:

[0063]

number

[0064] In the formula, y ijv is the CNR value of MRI scanner i (i∈{1,...,5}), subject j (j∈{1,...,128}), and SN-VTA voxel v (v∈{1,...,1807}), and α v is the mean CNR value for SN-VTA voxel v, X is the design matrix for the covariates of interest (age and sex), and β v is the vector of regression coefficients corresponding to X for SN-VTA voxel v, and γ iv and δ iv are the additive and multiplicative effects of MRI scanner i on SN-VTA voxel v, respectively, and ε ijk has mean 0 and variance

number

[0065] The ComBat harmonized CNR value can be defined as follows:

[0066]

number

[0067] In the formula, γ * iv and δ * iv are γ iv and δ iv is the empirical Bayes estimator of SN-VTA CNR. Harmonization of SN-VTA CNR can potentially be performed using pre-available computational packages (e.g., the MATLAB package hosted at https: / / github.com / Jfortin1 / ComBatHarmonization / tree / master / Matlab).

[0068] The harmonizer 358 can harmonize the identified imaging data sets and provide the harmonized data sets to a computing device (e.g., computing device 101, 201, physician device 103). In the above example, after processing by Combat, the final CH-CNR-NM voxels for each patient in the database can be analyzed with a variety of different algorithms. In some implementations, as described, the harmonized data sets can be further evaluated and / or displayed in any suitable manner.

[0069] Although the analysis device 305 is described as implementing each of the data manager and harmonizer via the processor 351, in other embodiments, an analysis device similar to the analysis device 305 can be composed of several instances of the units, components, and / or modules described above. For example, in some embodiments, a server can include several data managers, ML models, and / or harmonizers associated with one or more computing devices or computing devices. Furthermore, the terms data manager, ML model, and harmonizer are provided for illustrative purposes to describe the processes implemented, for example, by the processor 351. Thus, one or more of these modules can be combined into a single module or generally referred to as a processor configured to perform one or more processes or steps thereof. Furthermore, while the analysis device 305 is shown as a separate device from the computing device 201, in communication within such computing device 201, it can be understood that the analysis device 305 and the computing device 201 (or other analysis devices and / or computing devices described herein) can be implemented on one or more devices that include components (e.g., processors, memory, etc.) suitable for performing the processes described with reference to each.

[0070] Although analysis device 305 is described herein as including a data manager, an ML model, and a harmonizer, in other embodiments, analysis devices similar in structure and / or function to analysis device 305 can be configured such that some of the functions and / or modules described above can be performed on and / or by a computing device included in the system (e.g., computing device 201), for example, via a client-side application installed in the computing device (e.g., data handler 214 of FIG. 2). Similarly, in some cases, functions described as being performed on an analysis device (e.g., analysis device 305) can be performed on computing device 201, and vice versa.

[0071] In use, a computing device in the data harmonization system described herein (e.g., computing device 201 or any of the other computing devices described herein) can receive imaging data associated with a patient's region of interest (e.g., tissue or organ) acquired using an identified imaging device. The imaging data can be a biomarker-specific MRI dataset (e.g., an NM-MRI dataset). In some embodiments, the imaging data can be accompanied by or include supplemental data (e.g., metadata) indicative of clinical or historical data associated with the patient and / or data associated with a scanning device associated with acquisition of the imaging dataset. Alternatively, the computing device can receive the image data, and the computing device can be configured to provide an interface for adding supplemental information associated with the imaging data. The computing device can then transmit information associated with the imaging data and / or transmit the imaging data to an analysis device in the data harmonization system. In some cases, the computing device can also transmit information indicative of one or more sources of variability that may be of interest and / or one or more sources of undesirable variability. In some cases, the computing device can be configured to register (also referred to herein as importing) a scanning device (e.g., a new MRI scanner not registered in the data harmonization system). The registration data can include any suitable data related to the registered scanning device (e.g., manufacturer, model number, production, time, location, station, institution of use, etc.), all relevant data related to the subject (e.g., history, demographics, and other health related information related to the patient, study, etc.), and coefficients determined by the system to reduce variability resulting from or associated with the device, patient, etc.

[0072] FIG. 6 shows a flowchart illustrating an exemplary process of registering or acquiring a new scanner, according to some embodiments. In some cases, the computing device may indicate a reference data set that may be used to generate coefficients that, when applied, harmonize the identified imaging data set. The analysis device may receive the imaging data and supplemental data and generate a set of coefficients by implementing an algorithm (e.g., the ComBat algorithm). FIG. 7 shows a flowchart illustrating an exemplary process of generating coefficients by the analysis device, according to some embodiments. In some implementations, the computing device may check for predefined coefficients that may be applied to harmonize the identified imaging data set. The computing device may identify and indicate a predefined set of coefficients to be used to harmonize the identified imaging data set. For example, the computing device may indicate the identified scanning device that was used to generate the imaging data set. The identified device may be a registered scanning device associated with the data harmonization system, and the registration data is stored in a storage associated with the system. The registration data may include the registered scanning device and the coefficients determined by the system to reduce variability due to or associated with the device. The analysis device can recognize imaging equipment previously registered with the system and determine the identity of the imaging equipment (e.g., make, model, manufacturer, part number, etc.). The analysis device can determine and retrieve from storage associated with the system predetermined coefficients associated with the imaging equipment. The analysis device can then apply the coefficients to the identified imaging data set to calculate a harmonized imaging data set having reduced variability associated with undesirable sources of variability. FIG. 8 shows a flow chart illustrating an exemplary process of applying coefficients to an imaging data set to generate a harmonized data set, according to one embodiment. The analysis device can then transmit the harmonized imaging data set to the computing device.The computing device may suitably evaluate, display, and / or use (eg, via an interface) the harmonized imaging dataset, as desired.

[0073] In some embodiments, the data harmonization systems described herein (or components of such systems, such as any computing device described herein) can be configured to work with any imaging, including any suitable biomarker-specific MRI data.

[0074] Useful embodiments of imaging data harmonization methods include maintaining biologically meaningful variability (e.g., variability or differences associated with a diagnostic condition). When tailored to a specific biomarker, certain characteristics of the biomarker can be utilized to adapt the harmonization process that is best suited to that biomarker-specific imaging data. For example, neuromelanin (NM) is known to accumulate in dopaminergic midbrain neurons over a subject's lifetime, and thus NM-MRI signal can increase with age, providing biological variability that can be used to test NM-MRI harmonization approaches in healthy subjects. A further potential benefit of harmonization is the ability to improve both the reproducibility and statistical power of downstream analyses by removing unwanted non-biological variability. This can be particularly beneficial for biomarker development, where better reproducibility and statistical power can lead to more generalizable biomarkers.

[0075] FIG. 4 shows a flow chart describing an exemplary implementation of harmonizing imaging data using a data harmonization system 400, according to one embodiment. The data harmonization system 400 can be substantially similar in structure and / or function to the data harmonization systems described herein (e.g., system 100). The data harmonization system 400 is implemented using a cloud-based architecture having one or more portions implemented using a cloud-based computing system in a cloud-based computing framework. The system 400 includes a computing device 401 implementing a portal NM-101 (e.g., a web portal) configured to act as an interface to communicate with one or more cloud-based portions of the system 400 (e.g., an analysis device server, a database, storage, etc.). The computing device 401 of the system 400 communicates with a server system, collectively referred to as a VPC, via a communication network 406, which can be substantially similar to the communication network 106 described herein. The computing device 401 can be substantially similar in structure and / or function to the computing devices 101, 201 and / or analysis devices 105, 305 described above.

[0076] The cloud-based portion VPC of the system 400 includes an analysis device 405 having a portal server 402 and a processing server 403. The portal server 402 is configured to communicate with the computing device 401 via a NM-101 portal that serves as an interface for communication. The portal server 403 can receive instructions, commands, data indications, etc., sent from the computing device 401 via the interface, where the instructions indicate the imaging data set to be harmonized. The portal server 403 can receive metadata indicative of sources of variability that are undesirable and need to be reduced. For example, the portal server 402 can receive an indication of a scanning device used to generate the imaging data set. In some implementations, the portal server can recognize a scanning device that is not yet registered with the system 400. The portal server 402 can send instructions to the processing server to generate a coefficient set (e.g., by executing a ComBat algorithm) that, when applied to an imaging data set acquired using the scanning device, can be configured to reduce variability sourced from the scanning device and / or image acquisition parameters. The processing server 403 may generate the coefficients using any suitable process.

[0077] In an exemplary implementation, the processing server 403 may take up to one minute to complete generation of the coefficients. Once generated, the processing server 403 may send the coefficients and any other associated data to an NM-101 database 404, which may store the coefficients and associated data for future use, and may mark the scanning device as registered in the system 400. When the system 400 encounters new imaging data obtained from a new scanning device, the system 400 may be configured to automatically or semi-automatically (i.e., with user-provided input) generate new coefficients and store them while registering the scanning device (or any other suitable source of variability that may be desired to be reduced).

[0078] 4, initiating a registration process with the portal server 402 via the NM-101 portal may trigger a REST API implemented by the system 400 to initiate the Combat algorithm. The ComBat algorithm may be executed as follows: 1. Initiate a batch container that contains 51 reference voxel maps used to generate the default ComBat coefficients. 2. Initiate a parameter container containing the 51 reference ages and genders used to generate default combat coefficients. Note that the reference ages and genders align with the reference voxel map. The container can be any suitable executable computational component. For example, in some cases, the container can be a discrete computational environment. The environment can be set up within an operating system capable of running one or more applications. In some implementations, the container can include an entity (e.g., a CSV file ".csv" or a MAT file ".mat") that holds a dataset containing information related to the 51 reference ages and genders. 3. Query the NM-101 database for all cases matching the criteria {”scanner”=%TARGET_SCANNER%,”organization”=%TARGET_ORGANIZATION%}. For each case, multiple SN voxel maps can be computed and stored for concurrent execution. 4. Case by case... a. Ignore any voxel maps that have NAN values ​​in any of the voxels, calculate the average SN voxel map, and append it to the batch container. b. Append age and gender to the parameter container.

[0079] Run the Combat algorithm to generate new coefficients for the new scanner. Credit to https: / / github.com / Jfortin1 / ComBatHarmonization for the NeuroCombat algorithm in Python and Matlab.

[0080] The coefficients are stored in the NM-101 database and the coefficients are linked to the target scanner.

[0081] Once the coefficients are generated and stored, the analysis device 405 (ie, the portal server 402 or the processing server 403) can apply the coefficients to the imaging data to generate harmonized data.

[0082] In some implementations, the computing device 401 or the portal server 402 can recognize the indicated scanning equipment as a registered scanner with coefficients pre-generated by the system 400. The portal server 402 can then retrieve the coefficients and send a signal to a processing server 403, also implemented in the cloud-based framework VPC, to apply the coefficients.

[0083] As an example, prior to processing the SN voxel-by-voxel analysis, combat coefficients can be applied to the SN voxel map based on the recognized scanner used to capture the image. In some exemplary implementations, the system 400 can use the InstitutionName DICOM tag (0008, 1010) and StationName DICOM tag (0008, 0080) in the MRI image to identify the scanning institution and / or scanning device associated with the imaging data set. The system 400 can query the NM-101 database for combat coefficients used for the identified institution and / or scanning device. Appropriate coefficients can be obtained and applied to the SN voxels.

[0084] 5 is a flow chart illustrating a method for generating a harmonized data set using the data harmonization system, according to one embodiment. The method 500 can be implemented by a computing device that is similar in structure and / or function to the computing devices 101, 201, 401 and / or analysis devices 105, 305, 405 described above.

[0085] The method 500 includes, at 571, receiving a first dataset including a first plurality of images, the first dataset associated with a first source of magnetic resonance imaging (MRI). The first dataset can be a biomarker-specific MRI dataset, such as an NM-MRI dataset associated with a first patient and acquired using a first MRI scanner.

[0086] At 572, the method includes receiving a second dataset including a second plurality of images, the second dataset being associated with a second source of magnetic resonance imaging. The second dataset can be a biomarker-specific MRI dataset, such as an NM-MRI dataset associated with a second patient and / or acquired using a second MRI scanner.

[0087] At 573, the method 500 includes measuring a first level of neuromelanin based on one or more of the first plurality of images, the first level of neuromelanin associated with the first source. As an example, the first level of neuromelanin can include a CNR value associated with the first plurality of images.

[0088] At 574, the method 500 includes measuring a second level of neuromelanin based on one or more of the second plurality of images, the second level of neuromelanin associated with the second source. The second level of neuromelanin can include a CNR value associated with the second plurality of images.

[0089] At 575, the method 500 includes applying a first set of coefficients to the first level of neuromelanin, the first set of coefficients configured to adjust the first level of neuromelanin by reducing effects associated with the first source to generate a third level of neuromelanin. The first set of coefficients can be generated using a data harmonization system described herein (e.g., by implementing the ComBat algorithm described above). The third level of neuromelanin can be a ComBat harmonized (CH)CNR value associated with the first plurality of images generated as described herein.

[0090] At 576, the method 500 includes applying a second set of coefficients to the second level of neuromelanin, the first, second coefficients being configured to adjust the second level of neuromelanin by reducing effects associated with the first source to generate a fourth level of neuromelanin. The second set of coefficients can be generated using a data harmonization system described herein (e.g., by implementing the ComBat algorithm described above). The fourth level of neuromelanin can be a ComBat harmonized (CH)CNR value associated with the second plurality of images generated as described herein.

[0091] At 577, the method 500 includes generating a third harmonized data set including the third level of neuromelanin and the fourth level of neuromelanin.

[0092] FIG. 6 is a flow chart illustrating an example process 600 that may be implemented by a data harmonization system to capture an unregistered scanner, according to one embodiment. The process of FIG. 6 may be implemented by any suitable device (e.g., computing device, analysis device, etc.) in the data harmonization system described herein. A user, e.g., an administrator, may operate an interface (e.g., implemented by the interface manager 215) and navigate through the interface to check whether information associated as a social organization and / or scanner associated with the identified imaging data is available (exists). If so, the administrator may select an option and enter information linking the social organization and scanner identity to the imaging data and store the information in a database associated with the data harmonization system. If the scanner and / or social organization already exists as a registered entry in the system, the imaging data may be linked to the existing entry such that predefined coefficients associated with the entry may be used to harmonize the imaging data. Based on the selection of the scanner / social organization, the user may be prompted to initiate execution of an algorithm to generate coefficients and / or apply the coefficients to harmonize the identified imaging data.

[0093] Process 600 includes, at 601, a user logging into a portal (e.g., an NM-101 portal) as an administrator. At 602, the process includes the user navigating to a control tool, e.g., a page presented to users with administrator privileges called an "admin page." At 603, the process includes a decision point that evaluates whether a social organization associated with the scanner used to acquire the identified imaging dataset is already registered or exists in the data harmonization system. If the selection in response to the determination at 603 is "yes," process 600 includes, at 604, a decision point that evaluates whether a scanner used to acquire the identified imaging dataset is already registered or exists in the data harmonization system. If the selection at 604 is "yes," process 600 includes, at 606, a decision point that queries the user whether to run a data harmonization algorithm (e.g., a ComBat algorithm) on the identified imaging dataset based on the registration information associated with the social organization and the scanner. If the response to the query at 606 indicates "yes", the process continues to 605 where the process includes the user selecting an option to run the data harmonization algorithm. Alternatively, if the selection in response to the decision at 603 is "no", the process includes the user or administrator selecting a control tool at 608 for adding, for example in the form of an option, a new institution to the system's database as a social organization. Similarly, if the selection in response to the decision at 604 is "no", the process includes the user or administrator selecting, for example in the form of an option, a control tool at 609 (also reachable from 608) for adding, for example in the form of an option, a new scanner to the system's database. At 610, the user can add scanner details (e.g., from the name of the scanner reported in a DICOM tag if the imaging data is in DICOM format). At 611, the process includes the user saving the input regarding the scanner and / or social organization.At 612, the process includes submitting the input to the portal server to be stored in a database associated with the system. At 613, the process includes a decision point indicating the brand of the added scanner. As an example, if the scanner is a Siemens, the process includes at 614 saving the scanner with predefined or default coefficients associated with Siemens scanners. At 615, the process includes determining that the scanner is a Phillips and saving the scanner with predefined or default coefficients associated with Phillips scanners. At 604, the process includes determining that the scanner is a GE and saving the scanner with predefined or default coefficients associated with GE scanners. At 611, the process includes determining that the scanner is not of the identified manufacturer and saving the scanner without the default coefficients.

[0094] FIG. 7 is a flow chart illustrating an exemplary process 700 that may be implemented by a data harmonization system to initiate and implement the generation and storage of coefficients for harmonizing a data set, according to one embodiment. The process of FIG. 7 may be implemented by any suitable device (e.g., computing device, analysis device, etc.) in the data harmonization system described herein. The portion 701 of process 700 within the dashed box may be implemented in a computing device (e.g., computing device 101, 201, 401) that communicates via a portal with an analysis device (e.g., 105, 305, 405) as described herein with reference to systems 100 and 400 of FIGS. 1, 2, 3, and 4. As shown, a user may select a social organization and scanner associated with an imaging data set and initiate an algorithm (e.g., ComBat algorithm) via the portal to generate and / or apply coefficients for harmonizing the imaging data. A request may be received via the portal, and the system may search its database for existing and matching social organizations and scanners already registered with the system. If there are more than 10 matches, the system can load a default batch that can be used as a reference for the identified social organization or scanner. The default batch can be used to generate a voxel map that can be used to run the algorithm. Once the coefficients are generated, they can be stored in a database (i.e., update the information related to the identified social organization and / or scanner in the database). If there are not more than 10 cases, it is indicated that the system returns an error. In some cases, the system can be configured to perform other operations, such as adopting a predetermined reference map, etc.

[0095] Process 700 includes, at 701, a user initiating data harmonization by sending a request to run a data harmonization algorithm (e.g., ComBat algorithm). Portion 701 includes, at 701A, a user logging into a portal (e.g., NM-101 portal) and navigating to a control page at 701B. The process includes, at 701C, a user selecting a social organization and a scanner associated with the identified imaging data set to be harmonized, and, at 701D, initiating execution of the data harmonization algorithm (e.g., ComBat algorithm). At 701E, the process includes sending a request (e.g., from a computing device such as computing devices 101, 201, and / or 401) that can implement portion 701. Process 700 further includes, at 702, a query of all cases in a database associated with the system, the query configured to find a match between the social organization indicated (in the request of 701E) associated with the identified imaging data and a social organization stored in the system. The query also includes a search of all cases in a database associated with the system at 702 to find a match between the indicated scanner (in the request at 701E) associated with the identified imaging data and scanners stored in the system. Process 700 includes a decision point at 703 that evaluates whether more than a specified or threshold number of cases are found as hits in response to the query at 702. The threshold can be any desired number. In the example of FIG. 7, the evaluation at 703 is whether more than 10 cases are found as matches. If the answer is "no," the process returns an error at 707.Alternatively, if the response to the evaluation at 703 is "yes" (i.e., more than 10 cases were found as a match in response to the query at 702), the process includes loading, at 708, for each identified case, a default batch (e.g., a default batch of imaging data for that case) and parameter data associated with that case (e.g., default image acquisition parameters associated with the identified scanner and / or social organization for that batch of imaging data). At 709, the process includes processing the case (i.e., processing the default batch of imaging data sets based on the parameter data). The process includes, at 710, averaging all runs (imaging runs), each run including a set of voxels, into a single voxel map. At 711, the process includes appending the single voxel map to the batch of imaging data and appending one or more parameters associated with the identified imaging data set to the parameter data. At 712, the process includes evaluating whether all cases have been processed. If not, steps 709-711 are implemented for each case as described herein. If the selection in response to the evaluation at 712 is "yes", the process includes, at 704, running a data harmonization algorithm (the ComBat algorithm) on the identified imaging data set to generate coefficients associated with the scanner and social organization associated with the identified imaging data set. At 705, the process includes storing the coefficients, and at 706, the process includes updating information I in a database associated with the social organization to include the information and coefficients generated at 705.

[0096] 8 is a flow chart illustrating an example process 800 that may be implemented by a data harmonization system to apply coefficients to harmonize a dataset, according to one embodiment. Process 800 of FIG. 8 may be implemented by any suitable device (e.g., a computing device, an analysis device, etc.) in a data harmonization system described herein. As shown, the system may generate or obtain coefficients (as described herein) and apply them to the imaging data (e.g., an SN voxel map derived from multiple images). The system may perform a voxel-by-voxel analysis to generate a harmonized dataset.

[0097] Process 800 includes, at 801, sending a request to process MRI data (e.g., MRI imaging data), which may be biomarker-specific MRI imaging data, such as NM-MRI. The imaging data may include, for example, voxel values ​​associated with the SN or VTA regions of the brain obtained from the subject. The request may be sent from a computing device (e.g., computing device 101, 201, 401) that communicates via a portal with an analysis device (e.g., 105, 305, 405) as described herein with reference to systems 100 and 400 of Figs. 1, 2, 3, and 4, and with reference to process 600 of Fig. 6. At 802, process 800 includes executing an algorithm (e.g., NM-101 algorithm) for processing the data. Processing the data at 802 may include processes 600 and / or 700 described above. By way of example, this processing results in a voxel map associated with the imaging data. At 803, process 800 includes storing the raw voxel map in a database associated with the data harmonization system. For example, the voxel map can be an SN voxel map associated with the SN region of the brain and / or including an indication of a biomarker of interest (e.g., NM) as described herein. At 804, the process includes querying a database for coefficients associated with the scanner and social organization associated with the identified imaging data. At 805, the process includes evaluating whether suitable coefficients are present in the system's database. If the response to the evaluation at 805 is "yes" (e.g., positive identification and acquisition of suitable coefficients), the process includes at 806 applying the coefficients obtained from the database to the SN voxel map of the identified imaging data. The process includes at 807 completing the harmonization of the identified imaging data set at 809 by performing a SN voxel-wise analysis to generate a harmonized data set. If the response to the evaluation at 805 is "no," the process includes at 808 returning an error indicating that the identified imaging data set cannot be harmonized as requested.

[0098] Survey example A. Cross-scanner harmonization of neuromelanin-sensitive MRI for multi-site investigations An example embodiment of the systems and methods described herein was used for harmonization of NM-MRI data from three major MRI vendors (GE, Siemens, and Philips) with acquisition parameters that were not optimized across sites. The example study shows that the method successfully harmonized NM-MRI data while maintaining biologically meaningful variability (increase in NM-MRI signal with age), improving both the reproducibility and statistical power of the expected positive relationship between NM-MRI signal and age.

[0099] Dataset: NM-MRI and anatomical T1-weighted MRI data were collected in 128 healthy subjects from studies at Columbia University, The University of Amsterdam, and Stony Brook University. At Columbia University, 51 subjects (mean [SD] age: 34.5 [14.6] years, 22 women and 29 men) were collected on a 3T GE MR750 and 29 subjects (mean [SD] age: 29.0 [7.4] years, 15 women and 14 men) were collected on a 3T GE Signa Premier. At The University of Amsterdam, 12 subjects (mean [SD] age: 23.9 [4.1] years, 4 women and 8 men) were collected on a 3T Philips Ingenia and 12 subjects (mean [SD] age: 24.3 [2.1] years, 5 women and 7 men) were collected on a 3T Philips Ingenia Elition (hereafter referred to as Philips Elition). and Stony Brook University, where 24 subjects (mean [SD] age: 27.8 [8.9] years, 12 women and 12 men) were recruited at a 3T Siemens Prisma. None of the subjects had a history of neurological or psychiatric disorders as reported by the subjects. Due to differences in hardware and / or software, NM-MRI sequence parameters varied between the MRI scanners. A detailed description of the sequence parameters is provided in Table 1.

[0100] [Table 1]

[0101] Labels: CU, Columbia University; UA, University of Amsterdam; SBU, Stony Brook University; RO, readout direction; FOV, field of view; PE, phase encoding direction; SS, slice selection direction; TE, echo time; TR, repetition time; FA, flip angle; NEX, summation number; BE, bandwidth; MT, magnetization transfer pulse.

[0102] Data Processing: The data processing methods described herein can be implemented using a computing device, examples of which are described above with reference to Figures 1-3. NM-MRI data were preprocessed using a procedure pipeline combining Statistical Parametric Mapping (SPM) and Advanced Normalization Tools (ANT), which have previously been shown to achieve good test-retest reliability. The example procedure includes the following steps: (1) realignment to correct for motion using "SPM-Realign" if the averages were acquired separately (e.g., offline), and averaging of the realigned images using "SPM-ImCalc", (2) brain extraction of T1w images using "antsBrainExtraction.sh", (3) spatial normalization of brain-extracted T1w images to MNI152NLin2009cAsym space using "antsRegistrationSyN.sh" (rigid body + affine + deformable syn). (4) a step of coregistration of the NM-MRI images to the T1w images using “antsRegistrationSyN.sh” (rigid body); (5) a step of spatial normalization of the NM-MRI images to the MNI space defined according to the standard MNI template of the MRI scans by a single-step transformation combining the transformations estimated in steps (3) and (4) using “antsApplyTransforms”; and (6) a step of spatial smoothing of the spatially normalized NM-MRI images by a 1 mm full-width at half-maximum Gaussian kernel using “SPM-Smooth”.

[0103] The NM-MRI contrast ratio (CNR) maps were then estimated using the preprocessed NM-MRI images in the MNI template space. The NM-MRI CNR at each voxel is the ratio of the NM-MRI signal intensity at a given voxel (I V ) in the cerebral peduncle (an area of ​​white matter tracts known to have the lowest NM content) (I CC ) was calculated as the percentage of signal difference from

[0104] CNR V ={[I V -Mode(I CC )] / mode(I CC )}×100, where mode (I CC ) was calculated for each participant by fitting a kernel smoothing function to the histogram of all voxels in the CC mask.

[0105] CNR values ​​of 1,807 voxels within the substantia nigra-ventral tegmental area (SN-VTA) complex mask were harmonized using the ComBat harmonization model to remove non-biological variability (e.g., variability associated with fluctuations in the MRI scanner) while preserving biological variability associated with age and sex. This procedure is comparable to harmonization of outcome measures (e.g., SN-VTA CNR or regional brain volumes) instead of raw MRI signals. The computational model can be described as follows:

[0106]

number

[0107] In the formula, y ijv is the CNR value of MRI scanner i (i∈{1,...,5}), subject j (j∈{1,...,128}), and SN-VTA voxel v (v∈{1,...,1807}), and α v is the mean CNR value for SN-VTA voxel v, X is the design matrix for the covariates of interest (age and sex), and β v is the vector of regression coefficients corresponding to X for SN-VTA voxel v, and γ iv and δ iv are the additive and multiplicative effects of MRI scanner i on SN-VTA voxel v, respectively, and ε ijv has mean 0 and variance

number

[0108] The ComBat harmonized CNR value is defined as follows:

[0109]

number

[0110] During the ceremony,

number

number

[0111] Harmonization was also performed on raw MRI signals (e.g., NM-MRI signals in the CC and SN-VTA separately before calculating CNR), but could not remove non-biological variability (data not shown).

[0112] 9A and 9B, and 10A and 10B show the results of harmonization of imaging data acquired using various scanners. Specifically, the mean SN-VTA CNR distribution among subjects within each scanner was calculated and compared between scanners. As shown in FIG. 9A, the distribution of SN-VTA CNR values ​​obtained from imaging using five different scanners of three different types shows that the scanner effect (in the form of distribution shifts and scaling differences) is evident in the NM-MRI data. FIG. 10A shows the spread of the data in the distribution shown in FIG. 9A in the form of a box plot. The asterisks ( *) shows p<0.05 for an ANOVA comparing the means of median SN-VTA CNR distributions across MRI scanners. Variability in the data related to variations in scanners, acquisition parameters, etc. is visible by the spread of the means and standard deviations shown by the box plots in FIG. 10A. Each box plot is associated with each scanner listed (from left to right: GE MR750, GE Signa Premier, Siemens Prisma, Phillips Ingenia, and Phillips Elition). FIG. 9B shows the distribution of SN-VTA CNR values ​​obtained from imaging after data set harmonization. As shown, variability due to scanner effects, e.g., differences between scanners and / or acquisition parameters, etc., can be removed by data set harmonization, as shown by the normalized appearance of the distributions after harmonization, where all distributions have comparable means, standard deviations, scaling, etc. FIG. 10B shows the spread of the harmonized data in the distribution shown in FIG. 9B in the form of box plots, with each box plot associated with each scanner listed (from left to right: GE MR750, GE Signa Premier, Siemens Prisma, Phillips Ingenia, and Phillips Elition). The notation "ns" indicates that the data was not significant for the ANOVA comparing the median means of the SN-VTA CNR distributions across the MRI scanners. For the raw CNRs, the medians across the SN-VTA voxels were significantly different between scanners (F4,123=72.45, p<10-30, one-way ANOVA, FIG. 9A). For the harmonized CNRs, the differences were no longer evident, as indicated by comparable means, or means whose values ​​were closer to each other, for the data from the scanners (F4,123=0.13, p=0.969, one-way ANOVA, FIG. 9B).

[0113] Classification of MRI Scanners: The evaluation described herein can be implemented using a computing device, examples of which are described above with reference to Figures 1-3. Following the previous application of ComBat to the MRI data, the performance of the NM-MRI harmonization method was evaluated. The evaluation was performed by classifying the MRI scanners (GE MR750, GE Signa Premier, Philips Ingenia, Philips Elition, and Siemens Prisma) using a support vector machine (SVM) based on the patterns of SN-VTA CNR. Specifically, the input features for the SVM were the CNR values ​​of 1,807 voxels in the SN-VTA mask. Mean class classification accuracy was estimated using 5-fold cross-validation (1,000 iterations) of a 1-vs-1 error-correcting output sign linear SVM (MATLAB, fitcecoc with hyperparameter C optimized in an inner 3-fold cross-validation loop), which has previously shown improved performance over other multiclass classifiers, for raw CNR and ComBat-harmonized CNR individually. Given the imbalanced sample sizes across scanners, misclassification costs were set to be inversely proportional to the class frequencies per 1-vs-1 classifier. Statistical significance was determined using a permutation test in which the MRI scanner labels were randomly shuffled 10,000 times.

[0114] FIG. 11 shows classification results before and after data harmonization using a data harmonization system, according to one embodiment. FIG. 11 shows box plots of the average performance of MRI scanner classification using a 5-fold cross-validated linear SVM across 1,000 random folds of the data. The null distribution shows empirical chance-level performance (2.5-97.5th percentiles shown) determined by randomly shuffling each subject's MRI scanner label 10,000 times. Asterisks ( *) indicates p<0.05 for a permutation test comparing the mean classification accuracy over 1,000 random folds to a null distribution. For all box plots, the minimum, 25th, 50th (median), 75th, and maximum values ​​are shown. SVM classification results for the MRI scanners confirmed the differences between scanners (as shown to be far from the null distribution in the raw imaging data plotted on the left in Figure 11) and the ability to remove them via ComBat (as shown to be far from the null distribution in the harmonized CNR data plotted on the right in Figure 11). Similar to the univariate analysis for summary statistics (see Figures 9A, 9B, 10A, and 10B), multivoxel pattern analysis (MVPA) based on an SVM classifier using all voxel-wise raw CNRs within the SN-VTA mask showed a high chance accuracy rate for MRI scanner class classification above (mean ± standard deviation test accuracy rate = 86.5 ± 1.8%, 5-fold cross-validation; p = 0.0001, permutation test; Figure 11). With harmonized CNRs, the accuracy rate of the SVM classifier no longer exceeded chance (test accuracy rate = 29.5 ± 3.0%; p = 0.8542, permutation test; Figure 11).

[0115] B. NM-MRI harmonization preserves biologically meaningful variability To assess whether harmonization of NM-MRI data preserved inter-individual variability in biologically meaningful variables, SVR was used to predict age using all voxel-wise SN-VTA CNRs, as NM accumulates with age. The analyses described herein can be implemented using a computing device, examples of which are described above with reference to Figures 1-3.

[0116] Multivariate prediction of age: To determine the ability of the proposed NM-MRI harmonization method to preserve biologically meaningful variability, linear epsilon-insensitive support vector regression (SVR) was used to predict age from voxel-wise patterns of SN-VTA CNR. Specifically, input features for SVR were 1,807 voxel-wise CNR values ​​in the SN-VTA mask. The root mean square error (RMSE) and Pearson correlation coefficient (r) between actual and predicted age were estimated using 10-fold cross-validated (1,000 iterations) linear SVR (MATLAB, fitrlinear with hyperparameter C optimized in an inner 5-fold cross-validation loop) separately on raw CNR and ComBat-harmonized CNR. Statistical significance was determined using a permutation test in which the age values ​​of surrogate subjects were randomly shuffled 10,000 times. To determine the reproducibility of age prediction results, we selected 1,000 random subsets of subjects with varying sample sizes (32, 64, and 96) and calculated the Pearson correlation coefficient (r) between the SVR beta coefficients for each subset with the same sample size separately for the raw CNR and the ComBat-harmonized CNR. Due to the significant site effect observed in the raw CNR, site was deviated from the feature-by-feature regression trend before entering it into the SVR.

[0117] Standard univariate voxel-wise analysis of age effects: To determine the statistical power of the age effect (significant positive relationship with age), we performed a standard voxel-wise analysis via robust linear regression analysis (MATLAB, fitlm with “RobustOpts”, “on”) to predict CNR at all voxels v within the SN-VTA as follows:

[0118]

number

[0119] where n=5 represents the raw CNR (sex and four dummy variables for the five MRI scanners as confounding covariates) and n=1 represents the harmonized CNR (sex as confounding covariate). To correct for multiple comparisons, we defined the spatial extent of the effect as the number of voxels k, which showed a significant correlation with age (regression coefficient β at p<0.05). 1 voxel-level height threshold for t-test, one-tailed). Statistical significance was determined using a permutation test in which age values ​​were randomly shuffled 10,000 times as described above. Variation in effect size (k) was determined using a leave-one-out procedure in which the mean CNR of a given subject was extracted from voxels that had a significant effect within the remaining subjects.

[0120] The results showed that ComBat harmonization of NM-MRI SN-VTA CNR maintained biologically meaningful variability associated with age. SVR based on raw CNR was able to predict age significantly above chance because NM accumulates with age and this variability was maintained by the harmonization procedure. The results shown in Figures 12A and 12B show box plots of the average performance of age prediction using a 10-fold cross-validated linear SVR across 1000 random folds of the data for two performance metrics: root mean square error (RMSE, Figure 12A) and Pearson correlation coefficient (r, Figure 12B). The null distribution indicates empirical chance-level performance (2.5-97.5th percentiles shown) determined by randomly shuffling each subject's age 10,000 times. Asterisks ( *) indicates p<0.05 for permutation tests comparing mean age prediction performance over 1,000 random folds to a null distribution. ns: not significant for permutation tests comparing mean age prediction performance between raw CNR and harmonized CNR. For all box plots, minimum, 25th percentile, 50th percentile (median), 75th percentile, and maximum are shown. Note that the site effect on raw CNR before inputting it into the SVR deviated from the regression trend per feature. These results showed that the SVR based on raw CNR was able to predict age significantly above chance (mean ± standard deviation RMSE = 125.91 ± 3.83 months, 10-fold validation; p = 0.0002, Figure 12A, and permutation test; r = 0.51 ± 0.04; p = 0.0002, permutation test, Figure 12B). Performance was also significantly above chance based on harmonized CNR (RMSE = 118.35 ± 3.99 months, p = 0.0002, permutation test; Figure 12A, and r = 0.56 ± 0.05; p = 0.0002, permutation test; Figure 12B). This suggests that age-related biological variability remains in the data after harmonization. Importantly, there was no significant difference in age prediction using raw or harmonized CNR (ΔRMSE = 7.56 ± 5.53; p = 0.1551, permutation test; Δr = -0.06 ± 0.06; p = 0.7304, permutation test).

[0121] C. NM-MRI harmonization improves statistical power The effect of NM-MRI harmonization on statistical power was determined. Specifically, whether harmonization facilitates the detection of underlying age effects was examined using univariate voxel-wise analysis following our previous work. The analyses described herein can be implemented using computational devices, examples of which are described above with reference to Figures 1-3.

[0122] FIG. 13A shows a map of SN-VTA voxels whose raw NM-MRI CNR was positively correlated with age (thresholded at p<0.05, one-sided, voxel level) indicated by magenta pixels overlaid on the average NM-MRI CNR map from all subjects (NM-MRI template). 241 voxels out of 1,807 voxels showed an age effect. The middle image shows a map of an axial section of the subject's brain, and the top inset shows a coronal (left) and sagittal (right) section of the region. FIG. 13B shows a map of SN-VTA voxels whose harmonized NM-MRI CNR was positively correlated with age (thresholded at p<0.05, one-sided, voxel level) indicated by orange pixels overlaid on the NM-MRI template. 324 voxels out of 1,807 voxels showed an age effect. The middle image in Figure 13B also shows a map of the axial section of the subject's brain, and the top inset shows the coronal (left) and sagittal (right) sections of the region. In the raw CNR, a trend-level age effect in SN-VTA voxels was observed (241 of 1,807 voxels at p < 0.05, robust linear regression controlling for sex and MRI scanner, p 補正済み= 0.072, permutation test; Figure 13A). In the harmonized CNR, the magnitude of the age effect was significantly above chance level (324 voxels out of 1,807 voxels at p < 0.05, robust linear regression controlling for sex, p corrected = 0.042, permutation test; Figure 13B). A leave-one-out analysis (i.e., 128 folds) was performed to determine the unbiased correlation between age and the mean CNR at age voxels (voxels showing age effect). Here, the age voxels used to read out the mean CNR in a given subject were determined in the sample excluding that subject. The unbiased correlation of raw CNR was small but significant (r = 0.19, p = 0.031, partial Pearson correlation controlling for gender and MRI scanner), and this effect was numerically (but not statistically; z = 0.50, p = 0.62, z test) stronger for harmonized CNR (r = 0.25, p = 0.005, partial Pearson correlation controlling for gender). To further explore the effect of harmonization on statistical power, we compared the number of age-effect voxels per fold in the leave-one-out analysis between raw and harmonized CNR. Figure 14 shows histograms of the number of age-effect voxels across each of the 128 leave-one-out folds of raw and harmonized CNR. Asterisks ( * ) indicates p<0.05 for permutation tests comparing the mean number of age voxels from raw and harmonized CNRs. There were consistently fewer age effect voxels in raw CNRs (mean ± SD: 239.9 ± 16.4 voxels) compared to harmonized CNRs (323.9 ± 12.5 voxels), reaching statistical significance (difference = 84.0 ± 9.4 voxels, p = 0.002, permutation test; Figure 14). These results suggested that the effective statistical power was increased with harmonized CNRs using standard analytical procedures. These may be partially due to reduced degrees of freedom in the regression of raw and harmonized CNRs (the latter of which did not include site covariates).

[0123] D. Improved reproducibility through NM-MRI harmonization The replicability of the age effect, i.e., the ability to observe similar results across different sets of subjects, was investigated using a subsample of subjects from the combined dataset. The analyses described herein can be implemented using a computing device, examples of which are described above with reference to Figures 1-3.

[0124] In SVR, the effect of a variable (e.g., age) on prediction is coded into a β coefficient for each voxel, and the effect of interest is the multivoxel (spatial) pattern of β coefficients. To investigate the reliability of the pattern of β coefficients for predicting age, 128 subjects were randomly subsampled, and the similarity of the pattern was determined by calculating the spatial correlation between the SVR β coefficients from a given subset of subjects and those from all other random subsets of subjects with the same sample size. A significant interaction was found between sample size (n = 32, 64, 96, or 128) and CNR type (raw or harmonized), suggesting a significant improvement in the reproducibility of the harmonized CNR (F 3,7991992 = 9,288.75, p << 0.05, two-way ANOVA; Figure 15A). Figure 15A shows a bar graph depicting the mean (± SD) repeatability of β coefficients for 10-fold cross-validated linear SVR predicting age from 1,000 random 10-fold folds of data (n = 128) or from 1,000 random subsets of subjects (n = 32, 64, and 96) (left). Asterisks ( * ) indicates p<0.05 for two-way ANOVA comparing mean correlations of β coefficients between raw and harmonized CNRs for different sample sizes, or permutation tests for pairwise comparisons between raw and harmonized CNRs. Figure 15B shows histograms indicating the overlap of age-effect voxels across each of the 128 divisions of the LOO for raw and harmonized CNRs (right). Asterisks ( *) indicates p<0.05 for permutation tests comparing median Dice coefficients for raw and harmonized CNRs. Bonferroni post hoc tests revealed significant improvements in reproducibility of harmonized CNRs for all sample sizes except the smallest subsample. The sample sizes were 32 (raw CNR mean ± standard deviation: r = 0.34 ± 0.11, harmonized CNR: r = 0.34 ± 0.12; p = 0.817, permutation test; Figures 15A, 15B), 64 (raw CNR: r = 0.53 ± 0.08, harmonized CNR: r = 0.55 ± 0.07; p = 0.042, permutation test; Figures 15A, 15B), 96 (raw CNR: r = 0.73 ± 0.05, harmonized CNR: r = 0.75 ± 0.04; p = 0.006, permutation test; Figures 15A, 15B), and 128 (raw CNR: r = 0.99 ± 0.01, harmonized CNR: r = 0.99 ± 0.00; p = 0.002, permutation test; Figures 15A, 15B).

[0125] As a second test of the effect of harmonization on reproducibility, the reproducibility of age voxels was determined from univariate voxel-wise analysis. We calculated the overlap in spatial location of age voxels for each split of the leave-one-out analysis (measured using the Dice similarity coefficient (Dice, 1945)). We found that there was less overlap of age voxels in the raw CNR (median ± interquartile range: 0.956 ± 0.046) compared to the harmonized CNR (0.973 ± 0.022), and this difference (difference = -0.016 ± 0.035, p = 0.0002, permutation test; Figure 15A, 15B) was found to be statistically significant, further demonstrating improved reproducibility after harmonization.

[0126] E. Harmonization with ComBat Coefficients Using an Example System An embodiment of the system (e.g., NM-101 or other harmonization algorithm implemented using the computing device described above with reference to Figures 1-3) can be deployed out of the box using default Combat coefficients for GE, Siemens, and Philips brand scanners, while supporting the calculation of new coefficients for each scanner regardless of manufacturer brand. Some example use cases are provided below.

[0127] Onboarding of new equipment / scanners: The NM-101 algorithm can now be deployed with default combat coefficients calculated for GE, Siemens, and Philips scanners. The NM-101 algorithm was used to calculate voxel maps of SN per image for MRI images scanned from GE scanners (51 images), Siemens scanners (24 images), and Philips scanners (24 images). These voxel maps, along with the corresponding age and gender per image, are used to calculate default combat coefficients for GE, Siemens, and Philips scanners.

[0128] When a new social organization signs up for the NM-101 algorithm, an administrator can bring the social organization into the system through the NM-101 portal. Using administrator privileges, new social organizations can be added to the system and new scanners and their manufacturer models can be specified, which will be linked to the default combat coefficients in the database. [Table 2]

[0129] Combat Coefficient Generation: From the NM-101 Portal Admin page, an administrator can initiate the generation of new Combat Coefficients for any scanner registered in the system. The coefficient calculation can be run separately for each scanner and may take up to a minute to complete and update the database. Initiating the process in the NM-101 Portal triggers a REST API exposed by the NM-101 system that allows the Combat algorithm to be initiated and does the following: 1. Start a batch container that contains 51 reference voxel maps that will be used to generate the default combat coefficients. 2. Initiate a parameter container containing 51 reference ages and genders used to generate the default combat coefficients. Note that the reference ages and genders align with the reference voxel maps. 3. Query the NM-101 database for all cases matching the criteria {”scanner”=%TARGET_SCANNER%,”organization”=%TARGET_ORGANIZATION%}. For each case, compute multiple SN voxel maps and store them for concurrent execution. 4. Case by case... 5. Ignore any voxel maps that have NAN values ​​in any of the voxels, calculate the average SN voxel map, and append it to the batch container. 6. Append age and gender to the parameter container. 7. Run the Combat algorithm to generate new coefficients for the new scanner. Credit to https: / / github.com / Jfortin1 / ComBatHarmonization for the NeuroCombat algorithm in Python and Matlab. 8. Save the coefficients in the NM-101 database and link the coefficients to the target scanner.

[0130] Applying Coefficients: When processing a case through the NM-101 algorithm, combat coefficients can be applied to the SN voxel map based on the scanner used to capture the images before processing the SN voxel-by-voxel analysis. Using the InstitutionName DICOM tag (0008, 1010) and StationName DICOM tag (0008, 0080) for the MRI images for that case, the system can query the NM-101 database for the combat coefficients used for that case. These coefficients can be applied to the SN voxel map to adjust for batch effects from that scanner.

[0131] Data analysis can be performed before and after harmonization to evaluate the effectiveness of data harmonization. Case data from GE (51 cases), Philips (24 cases), and Siemens (24 cases) scanners were stored in the NM-101 database. The Combat algorithm was initiated and used to generate coefficients for the GE, Philips, and Siemens scanners, which were then stored back in the NM-101 database. These coefficients are treated as default coefficients for any scans captured with their corresponding scanner manufacturer brands (GE, Philips, Siemens).

[0132] Figure 16A shows density plots of SN CNR% (CNR-NM voxels at SN) for the GE, Siemens, and Philips cases evaluated. Figure 16B shows new density plots for the same data from the same cases as in Figure 16A after harmonization using Combat Coefficients obtained from the NM-101 database for each corresponding scanner (CNR-NM voxels at SN in Figure 17A).

[0133] F. Sample Size Analysis A study was conducted to observe the effect of sample size on the generation of Combat Coefficients for individual scanners. The analyses described herein can be implemented using a computing device, examples of which are described above with reference to Figures 1-3.

[0134] Combat coefficient sets were generated using batches of 5, 10, 15, 20, and 24 cases for each scanner type between Siemens and Phillips. Each coefficient set was then applied to the full set of cases for their corresponding scanner, and the data was harmonized against the reference GE dataset. The resulting SN CNR% density maps are plotted in Figures 17A and 17B, respectively.

[0135] To further demonstrate the improvement in Combat Coefficients when using larger batch sets, the resulting SN CNR% datasets are plotted as the difference from the reference GE density value. As the batch set size increases, the difference approaches zero, as shown in Figures 18A and 18B. For example, the dashed green line in Figure 18A (ref=0) or the dashed blue line in Figure 18B (ref=0) is the delta between that dataset and GE.

[0136] G. Application of coefficients generated using the ComBat algorithm to implement a fully automated workflow suitable for deployment in a hospital environment Challenges of Applying ComBat to a Hospital / Clinic Environment: The embodiments disclosed above describe a one-time application of the ComBat algorithm to an entire database to prepare the database for analysis by various algorithms, such as univariate and multivariate voxel-wise analysis. The following examples describe embodiments that include adaptations of the application of the ComBat algorithm to a hospital / clinic environment. The following examples described herein can be implemented using a computing device, examples of which are described above with reference to Figures 1-3.

[0137] In the aforementioned application, Combat was applied once to the clinical trial database at the end of data collection to create CH-CNR-NM voxels from CNR-NM voxels. This is possible because the data can be collected and then analyzed in its entirety at the conclusion of the clinical trial (or at a pre-specified midpoint) without any immediate modification. However, when the NM software program is deployed across hundreds or thousands of different clinics, there may be a constant daily influx of NM-MRI data. Rather than waiting months or years, as is common in large clinical trials, the NM data can be returned and reported to physicians, in some cases within hours of scanning. In this case, it may be impractical to run Combat on the entire database each time a new patient's NM-MRI is processed by the software.

[0138] Harmonization of NM-MRI data in a hospital environment: NM-MRI data is based on signal intensity, which can be highly dependent on the brand and model of the scanner, as well as the parameters that acquire each scan. This can be problematic for clinicians who may use automated workflows across multiple machines in hospitals where standard phantoms are not available. There can be multiple challenges that hinder the application of NM-MRI in the clinic. Three examples are presented here:

[0139] In a first example, a patient's neuromelanin concentration and volume may need to be measured over two separate time points. In this example, the patient undergoes two NM MRI scans on two different MRI machines with different parameters. These two scans have the potential for scanner-induced variability that confounds the ability to make meaningful comparisons between the two scans of either NM concentration and volume or symptom-specific voxels or disease-specific voxels.

[0140] In the second case where a patient undergoes an NM scan on an MRI machine and that NM scan must be compared to an NM scan of another patient measured on a different MRI machine, these two scans may have scanner-induced variability that confounds the ability to make meaningful comparisons between the two scans, either in NM concentration and volume, or in symptom-specific voxels, or disease-specific voxels.

[0141] Finally, in a third example, a standard control CNR-NM dataset needs to be generated and used, e.g., a reference dataset for a particular population or age group. To generate such a reference dataset, a database needs to be analyzed to generate standard control levels or standard voxel-wise patterns of neuromelanin concentration and volume. However, the database may contain CNR-NM MRI data acquired on a different MRI machine whose parameters are not optimized.

[0142] Application of ComBat Generation Factors: The disclosed embodiments include two solutions for harmonizing a patient's NM MRI data with Combat Generation Factors in a manner suitable for deployment in a fully automated clinical environment. These Combat Generation Factors are available on all NM MRI data going forward, eliminating the need to run Combat on further NM MRI data.

[0143] In an example implementation of the first solution, each MRI scanner is assigned a unique harmonization factor that aligns it with a standard reference MRI scanner. This harmonization factor is generated by acquiring several MRI scans with that scanner. These scans are stored in a database, and the Combat algorithm is run on the processed CNR-NM data stored in that database and on the CNR-NM data from the reference scanner database. This results in an MRI scanner-specific harmonization factor (called a scanner-specific factor) that is unique to each individual MRI scanner. The scanner-specific factor is then stored in a database of scanner factors. This process of generating scanner-specific factors for populating new MRI scanners for use with fully automated software is called scanner-specific harmonization and is detailed in Figures 6-8.

[0144] After the completion of the scanner-specific harmonization period, each time a new MRI is acquired with that scanner, the scanner-specific factors can be applied to the new CNR-NM data to harmonize it before reporting it to the physician or using it to perform statistical tests such as voxel-wise analysis methods (Figure 8). In this case, do not perform COMBAT on any new MRI scans after the scanner-specific harmonization period.

[0145] In an example implementation of the second solution, it may be possible to acquire NM-MRI data on a device for which scanner-specific coefficients do not exist, but it is not possible to run some test cases to harmonize the scanner and generate new scanner-specific coefficients. In these situations, the software can use alternative coefficients, called vendor-specific coefficients.

[0146] The vendor-specific coefficients can be generated by acquiring several scans on each vendor's MRI machine, where the scans may be acquired on different MRI machine models within a specific vendor family (e.g., different MRI models from Philips). Combat can be run on a database of CNR-NM voxels acquired from MRI machines of only one specific vendor, and the resulting data can be compared to CNR-NM voxels from a reference MRI vendor database (in this case, the GE database). The output from this comparison can be used to generate vendor-specific adjustment factors, called vendor-specific coefficients.

[0147] In the absence of scanner-specific coefficients, the vendor-specific coefficients can be applied to any new scans from the same vendor's MRI machine that have not yet undergone scanner-specific harmonization. In this case, Combat is not applied to any scans from the new scanner, and the coefficients take the place of the Combat implementation. This allows CNR-NM voxels to be harmonized to generate CH-CNR-NM voxels based solely on using the coefficients, without the need to run Combat directly on these unique voxels. One advantage of vendor-specific adjustment factors is that CNR-NM voxels can be harmonized to CH-CNR-NM even in the absence of scanner-specific adjustment factors. In an example implementation, Combat coefficients were generated from 5, 10, and 24 CNR-NM scans to determine the optimal number of MRI scans that can be used to generate accurate scanner-specific or vendor-specific adjustment factors. This is described in more detail in Figures 16-18 and the description associated with Figures 16-18.

[0148] Advantages: Because the Combat Coefficients give us CH-CNR-NM voxels, the following advantages also apply:

[0149] Voxel-wise analysis of a database of CH-CNR-NM voxels, which includes voxels measured from different MRI machines or the same MRI machine with different parameters, can result in increased power to detect voxel-based patterns than would be achieved using voxel-wise analysis on CNR voxels in the same database that would otherwise not have been subjected to harmonization. In one example, application of the voxel-wise analysis method to a database of CH-CNR voxels results in increased statistical power to detect voxels significantly associated with either a unique diagnosis (disease-specific voxels) or a unique symptom (symptom-specific voxels) than application of voxel-wise analysis to only CHR voxels.

[0150] Voxel-wise analysis of a database of CH-CNR-NM voxels, which includes voxels measured from different MRI machines or the same MRI machine with different parameters, can result in increased power to detect differences in neuromelanin concentration and volume than would be achieved using a voxel-wise analysis of CNR voxels in the same database that would not otherwise have been subjected to harmonization.

[0151] Voxel-wise analysis of a database of CH-CNR-NM voxels that includes voxels measured from different MRI machines or the same MRI machine with different parameters can result in improved reproducibility for detecting voxel-based patterns than would be achieved using voxel-wise analysis on CNR-NM voxels in the same database that would not otherwise have been subjected to harmonization. In one example, application of a voxel-wise analysis method to a database of CH-CNR-NM voxels results in increased statistical power for detecting voxels that are significantly associated with either a unique diagnosis (disease-specific voxels) or a unique symptom (symptom-specific voxels) than application of voxel-wise analysis to only CHR-NM voxels.

[0152] Voxel-wise analysis of a database of CH-CNR-NM voxels, which includes voxels measured from different MRI machines or the same MRI machine with different parameters, can result in improved reproducibility for detecting differences in neuromelanin concentration and volume than would be achieved using a voxel-wise analysis of CNR voxels in the same database that would not otherwise have been subjected to harmonization.

[0153] Multivariate voxel-wise analysis of a database of CH-CNR-NM voxels, including voxels measured on different MRI machines or on the same MRI machine with different parameters, can result in an improved ability to perform multivariate predictions than would be achieved using a voxel-wise analysis on CNR voxels in the same database that would not have otherwise been subjected to harmonization. In this example, age is... (Figure X).

[0154] In all of the embodiments listed above, the voxel-wise analysis can be univariate or multivariate.

[0155] The disclosed embodiments include methods for harmonizing NM-MRI data across sites and scanners to remove non-biological variability due to factors such as hardware and software differences. In addition to effectively removing non-biological variability, the harmonization method preserves biologically relevant variability (e.g., age effects) while increasing both reproducibility and statistical power. The disclosed systems and methods represent the first approach to harmonizing NM-MRI data and are an important step toward enabling and enhancing large-scale multi-site studies.

[0156] With implementation of the ComBat algorithm, we were successful in reducing systematic biases associated with scanners across multiple sites that did not fully harmonize acquisition protocols. We observed significant differences in SN-VTA CNR values ​​between two GE scanners (MR750 and Signa Premier), both at the same institute (Columbia University), but the MR750 data were acquired with a slice thickness of 3 mm, whereas the Signa Premier data were acquired with a slice thickness of 1.5 mm. ComBat's ability to harmonize these data suggested that the "batch" effect can encompass all sources of nonbiological variability and alleviate the need for fully matched acquisition protocols, which may not be possible due to software and hardware limitations.

[0157] In addition to removing non-biological variance from NM-MRI data, it can also be important to preserve biologically meaningful variability. The success of the harmonization method to facilitate multi-site investigations focused on biomarker development can include preserving relevant biological variability. In the case of NM-MRI biomarkers for the diagnosis of Parkinson's disease, schizophrenia, or any other neuropsychiatric condition, this can usually be an effect of diagnosis. Other biomarkers may instead focus on predicting treatment response or other clinically relevant outcomes. The disclosed embodiments and examples show proof of concept that ComBat can preserve biologically meaningful variability by showing that it preserves age-related variability among healthy individuals. The disclosed study represents, to the best of the inventors' knowledge, the first application of MVPA to NM-MRI data. Using the SVM method, we successfully classified MRI scanners from raw CNR and SN-VTA CNR of SVR, and successfully predicted age from SN-VTA CNR in both raw and harmonized CNR.

[0158] In conclusion, we disclose herein a system, device, and method for harmonizing NM-MRI data using ComBat, which effectively reduces nonbiological variability while maintaining biologically relevant variability, resulting in improved reproducibility and statistical power. ComBat harmonization is likely to benefit studies even when harmonizing acquisition parameters across various scanners prior to data collection, as is recommended for multisite studies. Importantly, however, current data suggest that harmonization is unlikely to be harmful or obscure biological effects under investigation, at least as long as they are identified in advance. This approach could pave the way for large-scale multisite studies to develop NM-MRI-based biomarkers.

[0159] Although various embodiments have been described as having particular combinations of features and / or components, other embodiments are possible that have any combination of features and / or components from any of the above described embodiments.

[0160] Some embodiments described herein relate to computer storage products with a non-transitory computer-readable medium (which may also be referred to as a non-transitory processor-readable medium) having instructions or computer code for performing various computer-implemented operations. The computer-readable medium (or processor-readable medium) is non-transitory in the sense that it does not itself include a transitory propagating signal (e.g., a propagating electromagnetic wave that carries information over a transmission medium such as space or cable). The medium and computer code (which may also be referred to as code) may be designed and constructed for a specific purpose or purposes. Examples of non-transitory computer-readable media include, but are not limited to, magnetic storage media such as hard disks, floppy disks, and magnetic tapes, optical storage media such as compact disks / digital video disks (CD / DVDs), compact disk read-only memory (CD-ROMs), and holographic devices, magneto-optical storage media such as optical disks, carrier wave signal processing modules, and hardware devices that are dedicatedly configured to store and execute program code such as application specific integrated circuits (ASICs), programmable logic devices (PLDs), read-only memory (ROM) devices, and random access memory (RAM) devices. Other embodiments described herein relate to computer program products, which may include, for example, the instructions and / or computer code described herein.

[0161] In this disclosure, unless expressly stated otherwise or clear from the context, reference to a singular article should be understood to include a plural article, and vice versa. Grammatical conjunctions are intended to express any and all disjunctive and conjunctive combinations of connected clauses, sentences, words, etc., unless stated otherwise or clear from the context. Thus, the term "or" should be understood to generally mean "and / or" and the like. The use of any and all examples or illustrative words (such as "for example," "such as," "including," etc.) presented herein is intended merely to facilitate understanding of the embodiments, and does not limit the scope of the embodiments or the claims.

[0162] Some embodiments and / or methods described herein may be performed by software (executed on hardware), hardware, or a combination thereof. Hardware modules may include, for example, general-purpose processors, field programmable gate arrays (FPGAs), and / or application-specific integrated circuits (ASICs). Software modules (executed on hardware) may be expressed in various software languages ​​(e.g., computer code), including C, C++, Java™, Ruby, Visual Basic™, and / or other object-oriented, procedural, or other programming languages ​​and development tools. Examples of computer code include, but are not limited to, microcode or microinstructions, machine instructions, such as those generated by a compiler, code used to generate web services, and files having higher level instructions executed by a compiler using an interpreter. For example, embodiments may be implemented using imperative programming languages ​​(e.g., C, Fortran, etc.), functional programming languages ​​(Haskell, Erlang, etc.), logic programming languages ​​(e.g., Prolog), object-oriented programming languages ​​(e.g., Java, C++, etc.), or other suitable programming languages ​​and / or development tools. Additional examples of computer code include, but are not limited to, control signals, encryption code, and compression code.

[0163] Also, various concepts may be embodied as one or more methods, examples of which are provided. The acts performed as part of a method may be ordered in any suitable manner. Thus, while shown as sequential acts in the exemplary embodiments, embodiments may be constructed in which acts are performed in an order different from that depicted, which may include performing some acts simultaneously.

Claims

1. 1. A method for harmonization of a neuromelanin dataset using vendor-specific Combat Coefficients, comprising: measuring a first level of neuromelanin associated with a first data set acquired using a set of MRI equipment manufactured by a first vendor; measuring a second level of neuromelanin associated with a second data set acquired using a reference MRI device manufactured by a second vendor different from the first vendor; implementing a machine learning (ComBat) algorithm for generating vendor-specific coefficients associated with the set of MRI machines, the vendor-specific coefficients being based on at least one of: (1) the first level of neuromelanin associated with the set of MRI machines; or (2) the second level of neuromelanin associated with the reference MRI machine; acquiring a third data set associated with a first MRI device manufactured by the first vendor; applying the vendor specific coefficients directly to the third data set from the first MRI machine manufactured by the first vendor to determine (1) the second level of neuromelanin associated with the reference MRI machine, or (2) an adjusted third level of neuromelanin associated with the first MRI machine compared to at least one of the second data sets; A method comprising:

2. 1. A method for assessing neuromelanin concentration in a region of interest in the brain of a subject, comprising: performing a neuromelanin magnetic resonance imaging (NM-MRI) scan on the subject; obtaining a neuromelanin data set from the NM-MRI scan; Optionally encrypting said neuromelanin data set; uploading the neuromelanin dataset to a remote server; Optionally decoding said data set; and (i) running a Combat algorithm to generate and apply coefficients to the entire data set, or (ii) applying scanner-specific coefficients to the data set, or (iii) applying vendor-specific coefficients to the data set, thereby harmonizing the data set; performing an analysis of the neuromelanin dataset, the analysis comprising: (i) comparing said neuromelanin dataset to one or more neuromelanin datasets previously obtained from said subject; (ii) comparing said neuromelanin dataset with a control dataset; (iii) comparing said neuromelanin dataset with one or more previously obtained neuromelanin datasets from different subjects; and generating a report including said neuromelanin analysis; uploading the report to a remote server; A method comprising:

3. 3. The method of claim 1 or 2, wherein the pre-combat NM dataset comprises contrast ratio-neuromelanin (CNR-NM) voxels.

4. 3. The method of claim 1 or 2, wherein the ComBat-adjusted NM dataset comprises ComBat-harmonized contrast ratio-neuromelanin (CH-CNR-NM) voxels.

5. 3. The method of claim 1 or 2, wherein the NM dataset comprises a measure indicative of the presence of neuromelanin in a midbrain region that contains a dopaminergic pathway in a subject.

6. The method according to claim 1 or 2, wherein the voxel-wise analysis procedure is performed on CH-CNR-NM voxels.

7. 1. An apparatus comprising a memory and a processor operatively coupled to the memory, The processor, receiving information associated with an imaging source; receiving a first dataset including a first plurality of images associated with the imaging source, each image in the plurality of images including a set of voxels; applying a mask to the first plurality of images to generate a voxel map, the voxel map being based on an average of values ​​associated with a subset of voxels from the set of voxels; calculating a first set of measures based on the voxel map, the first set of measures being based on indications of a presence of biomarkers in the plurality of images from which the voxel map was generated; providing the first measure set to a data harmonization algorithm configured to calculate a set of coefficients that, when applied to the first measure set, are configured to transform the first measure set into a harmonized second measure set, the second measure set having reduced variability associated with one or more identified sources of variability compared to the first measure set; receiving the second measure set and the set of coefficients associated with the transformation of the first measure set to the second measure set; storing the second set of measures and the set of coefficients in association with the information associated with the imaging source; The device is configured as follows.

8. 8. The apparatus of claim 7, wherein the processor is configured to calculate the first set of measures by determining a contrast-to-noise ratio (CNR) associated with each voxel from the subset of voxels, the CNR associated with each voxel being based on a difference between a first signal intensity associated with that voxel and a second signal intensity associated with a reference voxel, the reference voxel being known to contain a minimum content of the biomarker.

9. The device of claim 7 , wherein the biomarker is neuromelanin.

10. The apparatus of claim 7 , wherein the imaging source comprises at least one of an MRI scanner type, an MRI scanner, or a subject of an MRI scan.

11. 8. The apparatus of claim 7, wherein the coefficient set is such that the second set of measures has reduced variability associated with one or more identified sources of variability, including identified sources of additive and multiplicative variability.

12. 1. A non-transitory processor-readable medium storing code representing instructions for execution by a processor, the instructions including code that causes the processor to: Receiving a dataset comprising a plurality of images; receiving information associated with an imaging source associated with the data set; calculating a first set of measures based on the plurality of images; searching a database to determine a coefficient set associated with the imaging source, the coefficient set configured to transform the first measure set into a harmonized second measure set, the second measure set having reduced variability associated with one or more identified sources of variability compared to the first measure set; obtaining the set of coefficients associated with the imaging source; applying the set of coefficients to the first set of measures to generate the second set of measures having the reduced variability associated with the one or more identified sources of variability; storing the second set of measures associated with the set of coefficients and the information associated with the imaging source; The non-transitory processor-readable medium.

13. 13. The non-transitory processor-readable medium of claim 12, wherein the imaging source includes at least one of a brand of scanning equipment used to acquire the plurality of images, an identity of a scanning equipment used to acquire the plurality of images, or an object scanned to acquire the plurality of images.

14. The instructions include code that causes the processor to: determining, in response to the searching, that the coefficient set associated with the imaging source does not exist; and providing the first set of measures to a data harmonization algorithm, the data harmonization algorithm configured to generate the set of coefficients to convert the first set of measures to a harmonized second set of measures; The non-transitory processor-readable medium of claim 12 , further comprising:

15. The instructions include code that causes the processor to: applying a mask to the plurality of images, the mask being based on a region of interest; calculating an average voxel map associated with the region of interest based on the applying of the mask, the first set of measures being calculated based on the average voxel map; The non-transitory processor-readable medium of claim 12 , further comprising:

16. 16. The non-transitory processor-readable medium of claim 15, wherein the region of interest is the substantia nigra (SN) region.

17. The non-transitory processor-readable medium of claim 12 , wherein the plurality of images are acquired using biomarker-specific MRI.

18. 20. The non-transitory processor-readable medium of claim 17, wherein the biomarker is neuromelanin (NM).

Images