Method for imaging-based radiomics diagnosis of neuropsychiatric lupus based on machine learning

Abstract

The application provides a method for imaging-based diagnosis of neuropsychiatric lupus based on machine learning. The method comprises: obtaining a brain MRI image of a patient; segmenting the intracerebral region of interest of the preprocessed brain MRI image; extracting radiomics features in the region of interest; screening features related to the diagnosis and prediction of neuropsychiatric lupus; dividing the screened features into a first training set and a first test set; building N machine learning-based radiomics prediction models; saving the trained N machine learning-based radiomics prediction models; using multiple linear regression to save the radiomics prediction model corresponding to the optimal AUC brain region; obtaining the serological indicators of the NPSLE patient; dividing the serological indicators into a second training set and a second test set; building a joint prediction model combining radiomics features and serological indicators; and inputting the second test set into the joint prediction model to obtain the prediction classification result of the patient's neuropsychiatric lupus. The application fills the gap of magnetic resonance imaging in the diagnosis of NPSLE, and provides practical guidance for clinicians in assisting the diagnosis of NPSLE.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of neuropsychiatric lupus disease diagnosis, and particularly relates to a method for diagnosing neuropsychiatric lupus based on machine learning. BACKGROUND

[0002] Systemic lupus erythematosus (SLE) combined with neuropsychiatric (NP) disease is called neuropsychiatric lupus (NPSLE). NPSLE has a high incidence and mortality. In 1999, the American College of Rheumatology (ACR) proposed the definition of 19 kinds of NPSLE syndromes. A prospective multicenter study of 1206 SLE patients showed that 40.3% of patients had NP events, but most of them could not be attributed to SLE disease itself. However, NP events caused by SLE itself are more likely to be resolved by treatment, so it is very important to accurately distinguish NP events caused by SLE.

[0003] The pathophysiological mechanisms of NPSLE mainly include ischemic or autoimmune-mediated neuroinflammation. In clinical practice, the two pathophysiological processes may coexist at the same time. The imaging features of the two mechanisms are different and may overlap. It is difficult to obtain neuropathology, and the complexity of clinical manifestations, the overlap of non-lupus related NP events and other difficulties limit the optimization of NPSLE management and prognosis. Although different attribution models have been developed, however, to correctly attribute NP events to SLE, we still lack objective and specific biological and imaging markers.

[0004] Conventional magnetic resonance perfusion (cMRI) is the first choice of imaging technology for diagnosing NPSLE. The cMRI sequences include: T1 weighted image (T1WI), T2 weighted image (T2WI), fluid attenuated inversion recovery (Flair) and diffusion weighted imaging (DWI), but the abnormalities shown by these sequences are non-specific, and literature reports that 34%-42% of NPSLE patients appear normal on cMRI, so there may be certain limitations in diagnosing NPSLE based on the current cMRI sequence parameters. Although advanced neuroimaging techniques (fonctional MRI, fMRI), radionuclide brain scanning (SPECT or PET-CT), and dynamic susceptibility contrast enhancement (DSCMRI) largely overcome the limitations of cMRI by detecting hemodynamic and functional changes, but because of the high cost, shortage of equipment and related technical personnel, the popularity is still not enough. In addition, there are reports that the incidence of microbleeding in NPSLE is 13.5%, and traditional MRI and CT usually cannot detect microbleeding, and susceptibility weighted imaging (SWI) sequence is widely used in the detection of intracranial microbleeding, vascular malformations, Parkinson's disease and other aspects. Microbleeding is a challenge for clinicians in terms of treatment, especially in patients with thromboembolic events or aPL positive, who require long-term anticoagulant therapy. However, anti-platelet therapy and oral anticoagulant therapy are still controversial. The application of SWI in SLE patients is basically a technical blank. Considering the pathological changes of vasculitis in SLE patients, it is speculated that the SWI sequence may play a certain role in the detection of intracranial vascular abnormalities in SLE.

[0005] In summary, we urgently need a new method of imaging-based machine learning for diagnosing neuropsychiatric lupus, which aims to identify NPSLE through radiomics and multi-modalomics, and hopes to contribute to early and precise diagnosis and treatment strategies for NPSLE. SUMMARY

[0006] The present application aims to at least solve one of the problems in the prior art or related art.

[0007] To this end, the present application aims to provide a method of imaging-based machine learning for diagnosing neuropsychiatric lupus.

[0008] In order to achieve the above-mentioned purpose, the technical scheme of the present application provides a method for imaging-based diagnosis of neuropsychiatric lupus based on machine learning, which comprises the following steps: S1: obtaining a first data set; wherein the first data set is a brain MRI image of an NPSLE patient; the sequence of the brain MRI image includes T1WI, T2Flair, and SWI; S2: preprocessing the first data set; S3: automatically segmenting the intracerebral region of interest of the preprocessed first data set to divide the brain MRI image sequence of each NPSLE patient into N different brain regions; wherein N is a positive integer greater than 2; S4: extracting radiomics features in the intracerebral region of interest; S5: screening imaging features related to the diagnosis and prediction of neuropsychiatric lupus from the extracted radiomics features; S6: dividing the screened imaging features into a first training set and a first test set according to a predetermined proportion; S7: under different MR sequences, for N different brain regions, respectively building corresponding N machine learning-based radiomics prediction models; S8: inputting the first training set into the corresponding N machine learning-based radiomics prediction models for training, and saving the trained corresponding N machine learning-based radiomics prediction models; S9: inputting the first test set into the trained corresponding N machine learning-based radiomics prediction models, and using multiple linear regression to save the radiomics prediction model corresponding to the optimal AUC brain region; S10: obtaining a second data set; wherein the second data set is a serological index of an NPSLE patient; S11: dividing the serological index into a second training set and a second test set according to a predetermined proportion; S12: inputting the second training set into the radiomics prediction model corresponding to the optimal AUC brain region for training, and saving the trained radiomics prediction model corresponding to the optimal AUC brain region to build a joint prediction model combining radiomics features and serological indexes; S13: inputting the second test set into the joint prediction model to obtain a prediction classification result of whether the patient has neuropsychiatric lupus.

[0009] Preferably, the radiomics features in the intracerebral region of interest are extracted by the PyRadiomics library in the Python environment.

[0010] Preferably, the value of N is 8.

[0011] Preferably, the radiomics features include first-order features, shape-based features, texture features, and wavelet features.

[0012] Preferably, the radiomics features extracted are screened using a least absolute shrinkage and selection operator to select imaging features associated with the diagnosis prediction of neuropsychiatric lupus.

[0013] Preferably, the serological indicators of the NPSLE patient include serum anti-phospholipid antibodies, dsDNA antibodies, ribosome P protein antibodies, Ro-52 antibodies, immunoglobulins, and complements.

[0014] Preferably, the corresponding N machine learning-based radiomics prediction models are all random forests.

[0015] Advantages of the present application:

[0016] The method for diagnosing neuropsychiatric lupus based on machine learning and imageomics provided by the present application first segments the whole brain region of interest (ROI) of each NPSLE patient, then establishes a magnetic resonance radiomics model (i.e., the corresponding N machine learning-based radiomics prediction models) for each brain region based on the data features of traditional magnetic resonance and through a machine learning method. Further, the highest AUC value of the preferred imaging features is combined with the clinical features of the patient (i.e., the serological indicators of the NPSLE patient) to establish a joint prediction model combining the radiomics features and the serological indicators, and finally obtain the prediction classification result of the patient's neuropsychiatric lupus, filling the gap of magnetic resonance imageomics in the diagnosis of NPSLE.

[0017] Additional aspects and advantages of the present application will become apparent from the following description, or will be learned by practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A schematic flow chart of the method for diagnosing neuropsychiatric lupus based on machine learning and imageomics is shown. DETAILED DESCRIPTION

[0019] In order to more clearly understand the above-mentioned purposes, features and advantages of the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0020] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can also be implemented in other ways different from those described herein, therefore, the scope of protection of the present application is not limited to the specific embodiments disclosed below.

[0021] Figure 1A schematic flow chart of a method for diagnosing neuropsychiatric lupus based on machine learning and radiomics is shown. As shown in Figure 1 The method for diagnosing neuropsychiatric lupus based on machine learning and radiomics comprises:

[0022] Step S1: obtaining a first data set;

[0023] Step S2: preprocessing the first data set;

[0024] Step S3: automatically segmenting the intracerebral region of interest of the preprocessed first data set to divide the brain MRI image sequence of each NPSLE patient into N different brain regions;

[0025] Step S4: extracting radiomics features in the intracerebral region of interest;

[0026] Step S5: screening image features related to the diagnosis and prediction of neuropsychiatric lupus from the extracted radiomics features;

[0027] Step S6: dividing the screened image features into a first training set and a first test set according to a preset ratio;

[0028] Step S7: under different MR sequences, for N different brain regions, respectively building corresponding N machine learning-based radiomics prediction models;

[0029] Step S8: inputting the first training set into the corresponding N machine learning-based radiomics prediction models for training, and saving the trained corresponding N machine learning-based radiomics prediction models;

[0030] Step S9: inputting the first test set into the trained corresponding N machine learning-based radiomics prediction models, and using multiple linear regression to save the radiomics prediction model corresponding to the optimal AUC brain region;

[0031] Step S10: obtaining a second data set;

[0032] Step S11: dividing the serological indicators into a second training set and a second test set according to a preset ratio;

[0033] Step S12: inputting the second training set into the radiomics prediction model corresponding to the optimal AUC brain region for training, saving the trained radiomics prediction model corresponding to the optimal AUC brain region, to realize the building of a joint prediction model combining radiomics features and serological indicators;

[0034] Step S13: inputting the second test set into the joint prediction model to obtain the prediction classification result of whether the patient has neuropsychiatric lupus.

[0035] In the embodiment, the first data set is a brain MRI image of an NPSLE patient; and the sequence of the brain MRI image comprises T1WI, T2Flair and SWI.

[0036] In the embodiment, N is a positive integer greater than 2.

[0037] In the embodiment, the second data set is a serological index of the NPSLE patient.

[0038] In the embodiment, under different MR sequences, N machine learning-based radiomics prediction models are respectively established for N different brain regions; the different MR sequences comprise T1WI, T2Flair and SWI.

[0039] In the embodiment, the area under curve (AUC) is used.

[0040] In the embodiment, the method for diagnosing neuropsychiatric lupus based on machine learning and radiomics provided by the application is characterized in that: the whole brain region of interest (ROI) of each NPSLE patient is segmented for the first time, and then a magnetic resonance radiomics model (i.e., N machine learning-based radiomics prediction models) is established for each brain region based on the data characteristics of traditional magnetic resonance and through machine learning. Further, the preferred imaging features with the highest AUC value are combined with the clinical features (i.e., the serological index of the NPSLE patient) of the patient to establish a joint prediction model combining the radiomics features and the serological index, and finally the prediction classification result of the neuropsychiatric lupus of the patient is obtained, which fills the gap of magnetic resonance imaging in the diagnosis of NPSLE.

[0041] In an embodiment of the application, the radiomics features in the brain region of interest are extracted through the PyRadiomics library in the Python environment.

[0042] In an embodiment of the application, the value of N is 8.

[0043] In an embodiment of the application, the radiomics features include first-order features, shape-based features, texture features and wavelet features.

[0044] In an embodiment of the application, the least absolute shrinkage and selection operator (LASSO) is used to screen the imaging features related to the diagnosis and prediction of neuropsychiatric lupus from the extracted radiomics features.

[0045] In one embodiment of the present application, the serological indicators of the NPSLE patient include serum anti-phospholipid antibodies, dsDNA antibodies, ribosome P protein antibodies, Ro-52 antibodies, immunoglobulins, and complements.

[0046] In one embodiment of the present application, the corresponding N machine learning-based radiomics prediction models are all random forests.

[0047] The technical solution of the present application will be demonstrated below with one specific embodiment. The prospective analysis of MRI data of SLE patients in the present application is approved by the Human Experiment Ethics Standard Committee of the First Affiliated Hospital of China Medical University.

[0048] (1) SLE patients (including non-NPSLE and NPSLE patients) who visited the outpatient and inpatient departments of the First Affiliated Hospital of China Medical University were included. The diagnosis of SLE was based on the 1997 American College of Rheumatology (ACR) SLE diagnostic criteria or the 2012 International Lupus Cooperation Clinic Classification Standard; the diagnosis of NPSLE was based on the 1999 ACR naming classification of NPSLE. Common mild NP events in the general population were excluded: mild headache, mild anxiety, mild depression, and mild cognitive impairment.

[0049] Exclusion criteria: (a) severe metabolic-related cardiovascular and cerebrovascular complications; (b) history of drug or alcohol abuse; (c) head trauma or isolated cerebrovascular disease unrelated to NPSLE, (d) family history of mental illness, etc.

[0050] (2) Obtain the first data set; wherein the first data set is the brain MRI image of the NPSLE patient; the sequences of the brain MRI image include T1WI, T2Flair, and SWI. In this specific embodiment, T1WI is the abbreviation of T1 weighted imaging, SWI is the abbreviation of susceptibility weighted imaging, 3D-T1 is the thin layer T1WI sequence, and Sag cube flair is the thin layer T2Flair sequence.

[0051] The brain MRI image of the NPSLE patient was acquired using a nuclear magnetic resonance scanner 3.0T Pioneer GE MRI, and foam pads and earplugs were used during scanning to reduce the head motion of the subject and reduce the noise during scanning.

[0052] All subjects scanning parameters were as follows: (a) 3D-T1: TR 7.8 ms, TE 3.0 ms, FOV 24 cm x 24 cm, matrix size 240 x 240 mm2, slice thickness 1.0 mm, number of slices 176, no gap; (b) SWI: TR = Minimum, TE = out of phase, FA = 15°, FOV = 240 x 192 mm2, matrix size = 448 x 448, pixel size = 0.5 x 0.5 mm2, slice thickness = 2.0 mm, no gap, slices = 40. (c) Sag cube flair: TR = 6800 ms, TE = 100 ms, FOV = 240 x 216 mm2, matrix size = 240 x 220, pixel size = 1.0 x 1.1 mm2, slice thickness = 1.4 mm, no gap, slices = 130.

[0053] (3) Preprocessing the first data set. The first data set after preprocessing is segmented into the brain region of interest to divide each NPSLE patient's brain MRI image sequence into 8 different brain regions; we only draw 8 brain regions, in fact we have the data of the whole brain region, and we may extract other regions of interest in the later period, and only 8 brain regions are divided in this specific implementation.

[0054] The first data set was preprocessed, specifically including: using FreeSurfer (6.0) software to process 3D-T1 image data. First, cortical and subcortical brain tissue reconstruction was performed, including brain tissue resection, automatic Talairach transformation, normalization, subcortical white matter and gray matter volume segmentation, gray matter boundary segmentation, automatic topology correction, surface deformation to obtain the best gray matter / white matter / cerebrospinal fluid boundary. Then, the bilateral hippocampus, amygdala, choroid plexus, and lateral ventricle were taken as the region of interest (ROI). Sagcube flair and SWI were registered to the normalized T1 image using ANTs (i.e. Advanced Normalization Tools).(a) Individual 3D-T1 image and skull anatomy; (b) mapping relationship generated by linear and nonlinear transformation estimation; (c) deforming the sag cube flair and SWI images according to the above mapping relationship to complete registration. The regional segmentation and quality of each patient were manually checked by two experienced imaging physicians.

[0055] (4) Extract radiomic features in the brain region of interest. Radiomic features in the brain region of interest are extracted through the PyRadiomics library in the Python environment. Radiomic features include: first-order features, shape-based features, texture features, and wavelet features.

[0056] Among them, the first-order features describe the distribution characteristics of the signal intensity in the ROI. Shape-based features reflect the size and shape of the ROI in two-dimensional and three-dimensional space, regardless of the intensity distribution of gray scale. Texture features are based on the extraction of the following five matrices: (1) Gray Level Co-occurrence Matrix (GLCM); (2) Gray Level Size Zone Matrix (GLSZM); (3) Gray Level Run Length Matrix (GLRLM); (4) Neighboring Gray Level Difference Matrix (NGTDM); (5) Gray Level Dependence Matrix (GLDM). Wavelet features are extracted through wavelet decomposition of the image, covering features of different frequency bands.

[0057] According to the recommendation of the Pyradiomics development team, the initial settings for feature extraction include: 'binWidth' = 25; 'Interpolator' = sitk.sitkBSpline;'resampledPixelSpacing' = [1, 1, 1]; 'voxelArrayShift' = 1000; 'normalize' = True; 'normalizeScale' = 100.

[0058] (5) Select the imaging features related to the diagnosis and prediction of neuropsychiatric lupus from the extracted radiomic features. Use the least absolute shrinkage and selection operator to select the imaging features related to the diagnosis and prediction of neuropsychiatric lupus from the extracted radiomic features.

[0059] First, the least absolute shrinkage and selection operator (LASSO) is used to identify imaging features related to the diagnosis and prediction of neuropsychiatric lupus. LASSO is a linear regression containing L1 regularization, which is particularly useful for feature selection because it identifies the most relevant predictors from a large number of variables. The optimization problem of the lasso can be represented as:

[0060]

[0061] In the above formula, represents the observation value, represents the feature, represents the coefficient, is the regularization parameter, n represents the sample size, p represents the number of features, Represents the constant term or intercept term. LASSO modifies ordinary least squares regression (OLS) by adding a penalty equal to the absolute value of the coefficient, encouraging the model to shrink some coefficients to zero, thus performing feature selection.

[0062] (6) Divide the screened imaging features into a first training set and a first test set according to 7:3.

[0063] (7) Under different MR sequences (T1WI, T2Flair, SWI), for 8 different brain regions, build corresponding eight machine learning-based radiomics prediction models.

[0064] (8) Input the first training set into the corresponding eight machine learning-based radiomics prediction models for training, and save the trained corresponding eight machine learning-based radiomics prediction models.

[0065] (9) Input the first test set into the trained corresponding eight machine learning-based radiomics prediction models, and use multiple linear regression to save the optimal AUC brain region corresponding radiomics prediction model.

[0066] (10) Obtain a second data set. The second data set is the serological index of NPSLE patients; the serological index of SLE patients includes: serum anti-phospholipid antibody, dsDNA antibody, ribosome P protein antibody, Ro-52 antibody, immunoglobulin, and complement.

[0067] Antibody positivity is defined as: dsDNA antibody positivity when serum titer is higher than twice the threshold value; aPL positivity when at least one of anti-cardiolipin antibody (ACL), anti-β-2 glycoprotein 1 antibody and lupus anticoagulant test is positive within 12 weeks; and immunoglobulin includes IgG, IgA and IgM, and complement includes C3 and C4.

[0068] (11) Divide the serological index into a second training set and a second test set according to 7:3. Of course, the first training set and the second training set correspond to the same batch of patients, and the first test set and the second test set correspond to the same batch of patients.

[0069] (12) Input the second training set into the optimal AUC brain region corresponding radiomics prediction model for training, save the trained optimal AUC brain region corresponding radiomics prediction model, and build a joint prediction model combining radiomics features and serological indexes.

[0070] (13) Input the second test set into the joint prediction model to obtain the prediction classification result of whether the patient has neuropsychiatric lupus.

[0071] Further, we can also obtain the brain MRI images and serological indicators of SLE patients in the later stage, and accurately obtain the prediction classification result of whether the SLE patient has neuropsychiatric lupus according to the method of the present application. The NP event caused by the SLE patient is attributed to the SLE disease itself, and the NP event caused by the SLE itself is easier to be solved by treatment, so the present application brings practical guiding significance to assist the clinician in diagnosing NPSLE.

[0072] (14) Statistical analysis. Non-functional magnetic resonance imaging data were analyzed using SPSS 27 (IBM Software Analytics Corporation, Armonk, New York, USA). (a) Normal distribution of continuous variables was tested using the Kolmogorov-Smirnov D test. Data that met the normal distribution were expressed as mean (MD) ± standard deviation (SD), and data that did not meet the normal distribution were expressed as median (interquartile range). (b) Independent sample T test was used to compare normally distributed continuous variables between NPSLE and non-NPSLE. (c) Mann-whitneny U test was used to compare non-normally distributed continuous variables between NPSLE and non-NPSLE groups. (d) Chi-square test or Fisher's exact test was used to compare the classification variables of NPSLE and non-NPSLE groups.

[0073] The area under the receiver operating characteristic (ROC) curve (AUC) of the validation data set was analyzed and determined using the sklearn library in Python. AUC was used to compare the performance of different prediction scores. Linear plots were drawn using the Matplotlib library. P<0.05, the difference is a significant difference. Python (version 3.9.6; https: / / www.python.org / ) and R software (version 4.1.0; https: / / www.r-project.org) were used.

[0074] In summary, although the specificity of traditional brain MRI is not high, it is still the first choice of imaging examination method in clinic. MRI is highly popular, and all hospitals can achieve it. Moreover, the MRI equipment in each hospital is sufficient, and the relevant technical personnel do not need special training, as long as the MRI sequence number and parameters are adjusted. Therefore, in order to more effectively utilize the imaging features of traditional magnetic resonance and make up for the non-specificity of traditional brain magnetic resonance in NPSLE, the purpose of the present application is to divide the brain regions of NPSLE patients, extract and analyze the imaging features, and construct an imaging group model for NPSLE by using machine learning method. In addition, we will combine the preferred imaging model with the clinical characteristics of patients to establish a multi-modal group model, which can early and accurately judge whether SLE is complicated with central nervous system involvement, give effective management strategy and improve prognosis.

[0075] We have certain research results in the group model analysis of the 8 brain regions of the limbic system and choroid plexus of NPSLE patients. Compared with HC, the T1 sequence of the bilateral hippocampus of non-NPSLE and NPSLE patients showed higher AUC values (in the test set, the left hippocampus AUC of non-NPSLE group was 0.7305, the right hippocampus AUC was 0.7429; the left hippocampus AUC of NPSLE group was 0.69, and the right hippocampus AUC was 0.79). When comparing non-NPSLE and NPSLE groups, no difference and superiority of hippocampus was found, and in addition, we found that the T1 and SWI sequences of the left and right choroid plexus had the best prediction performance for the diagnosis of NPSLE (in the test set, the left choroid plexus T1-AUC was 0.7500, the left choroid plexus SWI-AUC was 0.7446; the right choroid plexus T1-AUC was 0.7681, and the right choroid plexus SWI-AUC was 0.7262).

[0076] It has been proved that the combined model of serum aPL (+) combined with bilateral choroid plexus SWI sequence further improves the prediction performance (in the test set, the left choroid plexus SWI-AUC was 0.7768, and the right choroid plexus SWI-AUC was 0.7578). Therefore, the method for imaging group diagnosis of neuropsychiatric lupus based on machine learning provided by the present application fills the blank of magnetic resonance imaging group in the diagnosis of NPSLE, and brings practical guiding significance for clinicians to assist in the diagnosis of NPSLE, and has an irreplaceable technical advantage in the intelligent diagnosis of NPSLE disease.

[0077] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for imaging omics diagnosis of neuropsychiatric lupus based on machine learning, characterized in that: include: Step S1: Acquire a first data set; wherein the first data set is a brain MRI image of a NPSLE patient; the sequence of the brain MRI image includes T1WI, T2Flair, and SWI; Step S2: preprocessing the first data set; Step S3: Automatically segmenting the brain region of interest on the preprocessed first data set to divide the brain MRI image sequence of each NPSLE patient into N different brain regions; where N is a positive integer greater than 2; Step S4: extracting radiomic features within the brain region of interest; Step S5: screening imaging features related to the diagnosis and prediction of neuropsychiatric lupus from the extracted radiomic features; Step S6: Dividing the screened imaging features into a first training set and a first test set according to a preset ratio; Step S7: Under different MR sequences, for N different brain regions, build corresponding N machine learning-based radiomics prediction models; Step S8: inputting the first training set into the corresponding N machine learning-based radiomics prediction models for training, and saving the trained corresponding N machine learning-based radiomics prediction models; Step S9: inputting the first test set into the corresponding N trained radiomics prediction models based on machine learning, and saving the radiomics prediction model corresponding to the optimal AUC brain region using multiple linear regression; Step S10: Acquire a second data set; wherein the second data set is serological indicators of NPSLE patients; Step S11: Dividing the serological indicators into a second training set and a second test set according to a preset ratio; Step S12: inputting the second training set into the radiomics prediction model corresponding to the optimal AUC brain region for training, and saving the trained radiomics prediction model corresponding to the optimal AUC brain region to build a joint prediction model combining radiomics features with serological indicators; Step S13: inputting the second test set into the joint prediction model to obtain a prediction classification result of whether the patient suffers from neuropsychiatric lupus.

2. The method for imaging omics diagnosis of neuropsychiatric lupus based on machine learning according to claim 1, characterized in that: The radiomics features of the brain region of interest were extracted using the PyRadiomics library in the Python environment.

3. The method for imaging omics diagnosis of neuropsychiatric lupus based on machine learning according to claim 1, characterized in that: The value of N is 8.

4. The method for imaging omics diagnosis of neuropsychiatric lupus based on machine learning according to claim 1, characterized in that: The radiomics features include: first-order features, shape-based features, texture features and wavelet features.

5. The method for imaging omics diagnosis of neuropsychiatric lupus based on machine learning according to claim 1, characterized in that: Least absolute shrinkage and selection operators were used to screen imaging features associated with the diagnosis and prediction of neuropsychiatric lupus from the extracted radiomic features.

6. The method for imaging omics diagnosis of neuropsychiatric lupus based on machine learning according to any one of claims 1 to 5, characterized in that: The serological indicators of the NPSLE patients include: serum antiphospholipid antibodies, dsDNA antibodies, ribosomal P protein antibodies, Ro-52 antibodies, immunoglobulins and complement.

7. The method for radiomics diagnosis of neuropsychiatric lupus based on machine learning according to any one of claims 1 to 5, wherein the corresponding N radiomics prediction models based on machine learning are all random forests.

Images