If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
1 Sijia Du, Cheng Yuan and Qinming Zhou contributed equally to this work.
Sijia Du
Footnotes
1 Sijia Du, Cheng Yuan and Qinming Zhou contributed equally to this work.
Affiliations
School of Biomedical Engineering, Shanghai Jiao Tong University, ChinaDepartment of Nuclear Medicine, Ruijin Hospital,Shanghai Jiao Tong University School of Medicine, China
1 Sijia Du, Cheng Yuan and Qinming Zhou contributed equally to this work.
Cheng Yuan
Footnotes
1 Sijia Du, Cheng Yuan and Qinming Zhou contributed equally to this work.
Affiliations
School of Biomedical Engineering, Shanghai Jiao Tong University, ChinaCollege of Medical Imaging, Shanghai University of Medicine and Health Sciences, China
School of Biomedical Engineering, Shanghai Jiao Tong University, ChinaDepartment of Nuclear Medicine, Ruijin Hospital,Shanghai Jiao Tong University School of Medicine, China
Department of Neurology and Institute of Neurology, Ruijin Hospital,Shanghai Jiao Tong University School of Medicine, ChinaCo-innovation Center of Neuroregeneration, Nantong University, China
We conducted experiments using multi-modality images of Chinese patients with MS.
•
We replaced manual lesion segmentation, which is time-consuming and has inter-rater variability with automatic lesion segmentation based on a deep learning network.
•
We explored the effect of radiomics on ARR prediction and verified that radiomics features extracted from PET/MR images could achieve a more accurate prognosis than traditional imaging indicators.
•
We provide a complete deep learning-based method to predict ARR from hybrid 18F-florbetapir PET and conventional MR (T1W and T2W) images of MS patients at their initial admission.
•
The workflow could serve as an automatic method in MS prognostic prediction and provide an insight into the MS relapse phenomenon from a radiomics point of view.
Abstract
Background Annualized Relapse Rate (ARR) is one of the most important indicators of disease progression in patients with Multiple Sclerosis (MS). However, imaging markers that can effectively predict ARR are currently unavailable. In this study, we developed a deep learning-based method for the automated extraction of radiomics features from Positron Emission Computed Tomography (PET) and Magnetic Resonance (MR) images to predict ARR in patients with MS.
Methods Twenty-five patients with a definite diagnosis of Relapsing-Remitting MS (RRMS) were enrolled in this study. We designed a multi-branch fully convolutional neural network to segment lesions from PET/MR images. After that, radiomics features were extracted from the obtained lesion volume of interest. Three feature selection methods were used to retain features highly correlated with ARR. We combined four classifiers with different feature selection methods to form twelve models for ARR classification. Finally, the model with the best performance was chosen.
Results Our network achieved precise automatic lesion segmentation with a Dice Similarity Coefficient (DSC) of 0.81 and a precision of 0.86. Radiomics features from lesions filtered by Recursive Feature Elimination (RFE) achieved the best performance in the Support Vector Machines (SVM) classifier. The classification model performance was best when radiomics from both PET and MR were combined to predict ARR, with high accuracy at 0.88 and Area Under the ROC curves (AUC) at 0.96, which outperformed MR or PET-based model and clinical indicators-based model.
Conclusion Our automatic segmentation masks can replace manual ones with excellent performance. Furthermore, the deep learning and PET/MR radiomics-based model in our research is an effective tool in assisting ARR classification of MS patients.
Multiple Sclerosis (MS) is a chronic autoimmune disease of the Central Nervous System (CNS), Approximately 2.8 million patients are suffering from MS worldwide (
Systematic literature review and network meta-analysis in highly active relapsing-remitting multiple sclerosis and rapidly evolving severe multiple sclerosis.
). Among all MS patients, Relapsing-Remitting Multiple Sclerosis (RRMS) is the most common clinical subtype accounting for approximately 85 of the affected population (
Systematic literature review and network meta-analysis in highly active relapsing-remitting multiple sclerosis and rapidly evolving severe multiple sclerosis.
Time-patterns of annualized relapse rates in randomized placebo-controlled clinical trials in relapsing multiple sclerosis: a systematic review and meta-analysis.
). Therefore, Annualized Relapse Rate (ARR) is an essential indicator for clinical prognosis, which could help physicians identify patients facing a higher risk of disease exacerbation.
Magnetic Resonance Imaging (MRI) is most commonly used for MS lesion assessment. However, the mid-term and long-term prognostic value of MRI for predicting disease evolution in MS patients remains debatable. Myelin Positron Emission Computed Tomography (PET) imaging using Aß tracers such as -PiB and -florbetapir has been used recently to explore the pathogenesis of demyelination of MS and showed a good correlation with the change in EDSS of RRMS patients in
), providing an insight into disease progression. Nevertheless, PET-derived quantitative parameters of demyelinated lesions demonstrated no predictive value for ARR of patients with MS in our previous study (
). Compared to conventional medical image parameters, radiomics describe more dimensional characteristics, including intensity, shape, and texture of images, and provide complementary information to clinical indicators. Furthermore, radiomics features extracted from multi-modality images could further capture within-patient heterogeneity, supporting the exploration of individualized disease characteristics associated with prognosis (
). Therefore, radiomics from multi-modality images such as PET and MRI may be a promising tool to explore ARR classification in patients with MS.
The relapse classification tasks based on radiomics are mainly performed in two ways: an end-to-end Convolutional Neural Network (CNN) or a workflow involving lesion segmentation and classification model estimation. In most CNN-based studies, researchers predict relapse directly from whole-brain images. Adrian Tousignant et al. presented an end-to-end CNN trained with multi-modality MRI for disease progression (
). Roca et al. fed age, sex, and Fluid-Attenuated Inversion Recovery (FLAIR) MRI into three predictors and took the weighted average score as a result (
). To evaluate the prognostic value for predicting disease in MRI, L Storelli et al. proposed a CNN to predict clinical and cognitive worsening with both EDSS and Symbol Digit Modalities Test (SDMT) (
). The methods mentioned above require a large amount of data to fully train the network, which is also less explainable than machine learning and radiomics-based methods.
With the introduction of radiomics, many researchers began to use machine learning with radiomics to make the most use of adequate information in medical images. Peng et al. extracted radiomics features at individual lesions level to predict their activity, which was fed into the SVM after being selected by the ReliefF algorithm (
). G Pontillo et al. combined radiomics feature, volumetric, and connectivity derived from MRI for EDSS prediction with segmentation masks acquired from Computational Anatomy Toolbox (CAT) software and formed a model with excellent intra- and inter-site generalizability (
The lesion masks used in these workflow are mainly obtained by manual annotation or existing segmentation software. Deep learning-based automatic segmentation is more efficient, reliable, and accurate to improve relapse classification performance.
proposed a modified 2D attention U-Net with a designed preprocessing scheme, reaching a DSC of 0.82. However, it only took the axial slices as input without considering information from the other two directions.
obtained a 3D lesion mask by integrating the 2D mask from three directions. However, the network is based on the Res-Net backbone, which resulted in a larger training cost.
In general, most of these studies focusing on MS disease course classification, whether employing CNN-based models or machine learning models, take EDSS as the relapse predictor and MR images as model input. Besides, these studies mentioned above didnt include PET and MRI for MS disease course classification. Inspired by the above progress, a workflow embedded in deep learning, that combines automatic lesion segmentation and ARR classification based on PET/MRI data, is needed to use the complementary information of PET and MRI.
2. Material and methods
2.1 Subjects
Twenty-five patients with a definite diagnosis of RRMS according to the 2017 revised McDonalds criteria were enrolled in this study from March 2019 to October 2020. All patients underwent clinical assessments, including clinical disability through the EDSS and -florbetapir PET/MR scan at their initial admission.
All patients followed up every 3–6 months. ARR of patients was calculated by the average number of relapses per year, within an 18-month follow-up after baseline PET/MR examination. The definition of relapses in patients with MS during follow-up was based on one of two conditions: a) clinical symptoms worsen or new MS-related symptoms appear; b) neuroimaging relapse, which is defined as newly developed hyperintensity lesions on T2WI or DWI at one follow-up scan (
) even if there are no significant clinical symptoms, and there must be an internal period of at least 30 days of symptom-free living between the new relapse and the previous one. All patients were divided into the group with relapse (ARR>0, n = 10) and the group without relapse (ARR = 0, n = 15). The clinical characteristics of the enrolled patients with or without relapse are summarized in Table 1 and listed in detail in Supplemental Table I.
Table 1The clinical characteristics of MS patients.
ARR = 0
ARR >0
p-value
Number
15
10
Age (year)
47.53± 13.63
35.40 ± 16.78
0.06
Gender (male/female)
9/6
4/6
0.33
Weight (kg)
63.67 ± 9.49
61.60 ± 11.05
0.62
EDSS
2.20 ± 1.44
2.90 ± 1.79
0.29
Disease duration (month)
65.80 ± 50.02
80.70 ± 125.90
0.68
Treatment
Methylprednisolone (n)
6
6
0.33
Teriflunomide (n)
1
0
0.41
No treatment (n)
8
4
0.51
EDSS: Expanded Disability Status Scale. ARR: Annualized Relapse Rate
5) Feature selection: selecting the most relevant features to ARR and removing highly correlated features.
6) ARR Classification: using the filtered features to form a predictive model. We will detail items 2–3 in Section 2.4 and items 4–6 in Section 2.5.
2.3 PET/MRI acquisition
A hybrid PET/MRI scan was performed with a Biograph mMR system (Siemens, Erlangen, Germany) with a NEMA PET resolution of 2.1 mm and 3T magnetic field. Syngo.via software (version 4.2, Siemens, Erlangen, Germany) was used for data registration, integration, and measurement of PET and MR images.. After intravenous injection of 287.9 ± 19.4 MBq of -florbetapir, dynamic PET acquisition in list mode over 60 min was started immediately. During PET acquisition, a 3D T1 magnetization-prepared rapid acquisition gradient echo (T1 MPRAGE, Repetition Time 1900 ms; Echo Time 2.44 ms; voxel size: 0.5 0.5 1.0 mm), a 3D T2-weighted fluid-attenuated inversion recovery (T2 FLAIR, Repetition Time 5000 ms; Echo Time 385 ms; voxel size: 0.5 0.5 0.9 mm) were carried out using the head coil.. The PET image was reconstructed by a point spread function algorithm with 344 344 pixels, 4 iterations, 21 subsets, and a filter with a full width at half maximum of 2 mm.
) was applied at the voxel level on PET scans to produce a parametric map of -florbetapir binding measured as the Distribution Volume Ratio (DVR, defined as the ratio of the total distribution volume between the target and the reference region). The average DVR in Damaged White Matter (DWM) lesions for each patient was calculated. Additionally, the total volume of T2 lesions as an indicator of macroscopic DWM lesion load was calculated for each patient.
2.4 Automatic lesion segmentation
A 2D network was designed for automatic lesion segmentation using T2WI and PET. The network structure is shown in Fig. 2. It has a backbone of U-Net (
), containing a multi-branch encoding path and a decoding path. The encoding path consists of five convolution-pooling layers to extract multi-resolution features from these modalities. To handle the multi-modality and multi-resolution data obtained above, we adopted a Multi-Modal Feature Fusion block (MMFF) and Multi-Scale Feature Upsampling block (MSFU) proposed by
. The MSFU module contains two convolution layers with kernel sizes of one and three to extract advanced features further. The feature maps of different modalities from the MMFF module were integrated by channel-wise concatenation. After that, features of different scales were input into the MSFU module to obtain a segmentation mask of the same size as the input images by upsampling. The proposed network structure could utilize MR and PET complementary information to achieve accurate automatic segmentation.
Fig. 2The topology of the proposed automatic segmentation network.
To fully use the three-dimensional information in PET/MR data, networks were trained and selected with slices at transverse, coronal, and sagittal axes. The automatic segmentation masks of twenty-five patients in the ARR classification task were obtained using the optimal network in each direction and integrated into the resulting masks with a majority voting algorithm.
To verify the superiority of our lesion segmentation network, we compared our automatic segmentation masks with those obtained in the Lesion Segmentation Tool (LST) toolbox, an open-source toolbox originally developed for MS lesions segmentation (
). Evaluation indexes, including Dice Similarity Coefficient (DSC), precision, sensitivity, Hausdorff Distance (HD), and Average Symmetric Surface Distance (AD), were firstly calculated on the scale of each patient. The final evaluation of our auto segmentation network was calculated by averaging the segmentation performance of twenty-five patients in the ARR classification task.
2.5 Classification model establishment
In this section, classification models were established after feature extraction and selection to predicted whether ARR is equal to zero or greater than zero.
Initially, radiomics features in T1WI, T2WI, and PET were extracted from the automatic segmentation lesion masks based on Pyradiomics. Various well-designed features were extracted from original and filtered images. Eventually, there was the computation of approximately 4278 features with the analysis methods available in Pyradiomics.
The extracted high-dimensional radiomics features were selected to filter out those highly related to ARR. We tested the different combinations of feature selection methods and classifiers for ARR classification. Three feature selection methods were considered in our study, including Recursive Feature Elimination (RFE), ReliefF, and Least Absolute Shrinkage and Selection Operator (LASSO). The classifiers we used contained Support Vector Machine (SVM), Random forest (RF), K-Nearest Neighbors (KNN), and Logistic Regression (LR). As a result, twelve models to be tested were formed. We compared these models through comparative experiments. The best performing network took selected features of T2WI, T1WI and PET as the input to predict the level of patients’ ARR using the complementary information between MR and PET.Selected features in three modalities were fed into the classifier to form a model that could predict the level of patients’ ARR using the complementary information between modalities.
Because the correlation between features was not considered in the algorithms mentioned above, there were still many features with high correlation in the robust feature subset obtained before. Redundant features should be deleted by calculating the Pearson product-moment correlation coefficient (Pearson’s r). In all pair features with a higher correlation coefficient (r>0.85), the one with the lower Area Under the Curve (AUC) when predicting ARR will be removed.
For a specific combination of a feature selection method and a classifier, the following process was conducted. The features extracted by Pyradiomics were normalized first before being fed into the classification model. After that, the feature selection method was used to filter out features with a high correlation to ARR. The redundant features were deleted to ensure the correlation coefficients between the rest of the features were lower than 0.85. Finally, the filtered features were fed into the classifier to evaluate the classification results.
As a result, the best-performing feature selection method and its corresponding classifier were retained to build an accurate model for annualized relapse rate classification by comparing the evaluation indexes of twelve models. Our python code for the algorithm mentioned above can be found in the link as https://github.com/Scarlett213/ARR-classification-model-in-MS.
3. Results
3.1 Lesion segmentation
3.1.1 Methods comparison
The performance of our lesion segmentation network and the LST are shown in Table 2, which shows that our automatic segmentation network outperforms LST in all aspects and significantly reduces HD and AD.
Table 2Comparison of segmentation results between our method and LST. The higher DSC, precision, and sensitivity mean better performance, and the other parameters are vice versa.
We visualize the lesion mask achieved by different methods from the axial, sagittal, and coronal views in Fig. 3. Our segmentation network achieved better performance in small lesions segmentation compared with LST. Besides, the automatic segmentation mask has a more accurate and distinct boundary which results in a significant improvement in HD and AD. Additionally, the segmentation network achieves robust performance in patients with both large and small lesions. Compared with the LST the proposed network can provide a more accurate lesion mask to ensure the robustness and reliability of extracted radiomics features, improving the classification accuracy.
Fig. 3Visual comparison of lesion segmentation performance. The red area represents the lesion mask obtained by methods noted below, and the white arrows indicate the location of the lesions where our segmentation network achieved more accurate results compared to LST. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
In order to verify the influence of each module on the segmentation mask, ablation study were carried out. The evaluation indicators are the same as above. The results in Table 3 show that both the multimodal fusion module and the majority voting algorithm can improve the performance of the segmentation algorithm. Here, multimodal fusion module contains both the Modal Feature Fusion block (MMFF) and Multi-Scale Feature Upsampling block (MSFU) to fuse information from MR and PET. And the majority voting algorithm was used to take advantage of the three-dimensional information of MR and PET at transverse, coronal, and sagittal axes. Our network improved the classification performance of both AD and HD, which means the multimodal fusion module and the majority voting algorithm can assist the segmentation network to better identify the lesion boundary.
Table 3The impact of each module on the segmentation algorithm.
3.2 Comparison of different features selection and classifier methods for predicting ARR
From the results in Table 4, PET/MR multi-modality radiomics model, whose features were filtered by RFE, achieved the best performance in the SVM classifier with the highest F1 score, AUC, accuracy, and a good balance between recall and precision. The classification results change significantly with the type of classifier. So contrast experiences are necessary to find the most suitable classifier for a specific feature selection method.
Table 4Performance evaluation of different combinations of feature selection methods and classifiers for ARR classification. Num refers to number of filtered features. P, R, F1, SP, ACC and AUC are the abbreviations for precision, recall, F1-score, specificity, accuracy and area under the ROC curves, respectively.P, R and SP are the abbreviations for precision, recall and Specificity, respectively.
Classifiers
SVM
RF
KNN
LR
Methods
Num
P
R
F1
SP
ACC
AUC
P
R
F1
SP
ACC
AUC
P
R
F1
SP
ACC
AUC
P
R
F1
SP
ACC
AUC
LASSO
145
0.27
0.30
0.29
0.47
0.40
0.50
0.67
0.20
0.31
0.93
0.64
0.43
0.39
0.90
0.55
0.07
0.40
0.47
0.50
0.50
0.50
0.67
0.60
0.65
RFE
77
0.89
0.80
0.84
0.93
0.88
0.96
0.71
0.50
0.59
0.87
0.72
0.71
0.41
0.90
0.56
0.13
0.44
0.76
0.86
0.60
0.71
0.93
0.80
0.96
ReliefF
202
0.33
0.20
0.25
0.73
0.52
0.48
0.38
0.30
0.33
0.67
0.52
0.73
0.42
0.80
0.55
0.27
0.48
0.62
0.43
0.30
0.35
0.73
0.56
0.62
SVM: Support Vector Machines, RF: Random forest, KNN: K-Nearest Neighbors LR: Logistic Regression, LASSO: Least Absolute Shrinkage and Selection Operator, RFE: Recursive Feature Elimination.
To further evaluate the performance of our ARR classification model, we compared the optimal model achieved above with the proposed ones in research focusing on MS prognosis prediction based on radiomics. The radiomics features in compared methods were extracted from MR images with the manual segmentation mask and fed into the prediction model used in the papers. Results were also evaluated by indexes calculated from five-fold cross-validation and listed in Table 5.
Table 5Comparison of our proposed pipeline with other methods.
Our proposed model showed better classification performance with a clear-cut improvement in precision and accuracy compared with previous MR-radiomics models. It shows that our model, which predicts ARR from PET/MR radiomics features, outperforms these methods using only MR images. Besides, features extracted from automatic segmentation masks can also form a model with a good performance. In general, our multi-modality deep learning-based pipeline has promising potential in ARR classification, which can offer an accurate result and reduce the time cost of lesion delineation.
3.3 Classification model comparison
We additionally built three models based on single modality images to verify the necessity to use multi-modality data. We used the radiomics features extracted from the lesions in each modality to form a model separately according to the methods mentioned in Section 2.5. and evaluated the contribution of a single modality with both the evaluation indexes and the mean ROC curve acquired by five-fold cross-validation. The results of twelve combinations of feature selection methods and classifiers for a single modality are in Supplemental Table III in detail.
We also built a classification model with clinical indicators to reveal radiomics features’ superiority, which includes the uptake in the DVR maps, lesion load in T2WI, patients’ age, gender, EDSS, and disease course. Similarly, clinical indicators were fed into four classifiers to choose the most suitable prediction method for these features. See Supplemental Table IV for detailed comparison results.
The results of five different models for ARR classification are shown in Table 6. Among the three single modalities, T2WI achieved the best performance. The classification results are weaker to predict ARR only using PET, and it has a poor performance to predict ARR with T1WI only. The evaluation indexes significantly improved with the combination of three modalities, which revealed the necessity of using multi-modal imaging in the clinical diagnosis of MS. Besides, the clinical indicator-based model is still inferior to radiomics-based models. Disease courses and EDSS are indicators that need to be calculated during follow-up. By comparison, the model established on the radiomics features extracted from patients’ initial admission images is not only better than the model based on clinical indicators in accuracy; but also makes it possible to learn patients’ ARR level as early as possible to shift more attention to patients with higher ARR in time.
Table 6The evaluation indexes of different modalities for ARR classification in five-fold cross-validation.
Model
Num
Precision
Recall
F1
Specificity
T1WI
100
0.50
0.40
0.44
0.73
T2WI
96
0.83
0.50
0.62
0.93
PET
40
0.50
0.50
0.50
0.67
CI
6
0.62
0.80
0.70
0.67
All Modality
77
0.89
0.80
0.84
0.93
CI: clinical indicators, Num: the number of features, F1: F1 score.
We plot the ROC curve of the radiomics-based model and the clinical indicator-based model in Fig. 4. These models have better predictive performance than PET DVR (AUC=0.42) and T2 load (AUC=0.37).
Fig. 4The ROC curve of different models in ARR classification. The red dotted line represents the ROC curve of random classification. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
Representative baseline PET/MR and follow-up MR images of relapse-free (Patients No.23, male, 28y, EDSS 2.5, treated by methylprednisolone) versus relapsed patients (Patients No.1, female, 19y, EDSS 2.0, treated by methylprednisolone) are presented in Fig. 5a. Although the baseline EDSS and treatment were similar in the two young patients, their ARR within an 18-month follow-up was different. To explore the discrepancy in their PET/MR images at initial admission, we analyzed the specific value of their radiomics features in our classification mode and visualized five features with the largest numerical difference in each modality in Fig. 5b–d. Since the distribution range of each feature is different, we normalized them to the interval between 0 and 1 before visualization. The names and meanings of each feature are listed in detail in Supplemental Table II.
Fig. 5Comparison of two patients. Baseline PET-DVR map, baseline, and follow-up T2W images from left to right of two representative patients. (b-d) Top five features of PET, T2WI, and T1WI radiomics that differed most between patients with ARR equal to 0 and ARR greater than 0. The specific meanings of the features are listed in Supplemental Table II.
Our study proposed a workflow that started with automatic segmentation, proceeded to feature extraction and selection, and ended with ARR classification. In such a pipeline, we incorporated PET as a new modality into MS research to verify the superior performance of radiomics features compared with conventional images or clinical indicators. Our study also demonstrated the need of combining PET and MR radiomics for ARR classification, which could complement each other to improve the accuracy.
To the best of our knowledge, it is the first study to use ARR as a target indicator of MS disease progression, offering new insight into relapse. The result shows that our model can successfully distinguish patients with similar EDSS but dissimilar ARR, which may serve as a purposeful tool to differentiate patients at different risks of relapse at their initial admission. We verified in this thesis that the PET/MR multi-modality model reaches a higher classification precision and accuracy, improving the AUC significantly compared with the MR-based model (
). In addition, our results also show that the PET/MR multi-modality model not only outperforms the single-modality ones but also achieves obvious advantages compared with the clinical indicators-based model.
The workflow embedding automatic segmentation mask and PET/MR images outperforms these methods extracting radiomics features in MRI through manually segmented lesions. The proposed method achieved the highest F1 score, AUC, accuracy, and a good balance between recall and precision. In terms of automatic segmentation performance, our U-Net-like lesion segmentation network significantly reduces the HD and AD of lesion masks by major voting algorithms with an HD of 3.46mm and an AD of 0.53mm The accurate and reliable 3D lesions masks guarantee the reliability of classification. Besides, it has a concise network structure to improve the training efficiency while ensuring the accuracy of the lesion segmentation mask compared with the ResNet-based method15.
We believe that this improvement in predictive performance may be associated with the information on specific myelin changes reflected by PET, which is complementary to conventional MRI information. Among all these selected radiomics features for predicting the relapse of MS patients, the top 5 features from T1W images showed that the initial images of patients with relapse within 18-month follow-up had a smaller high gray-level zone. The features from T2WI, such as GLCM, indicated that the local grey level homogeneity and the neighboring intensity were higher for patients with relapse. The radiomics features from both T1W and T2W images suggested that the patients with high levels of inflammatory and signal heterogeneity were more likely to have disease relapse during follow-up. Besides, the features from PET, such as GLSZM, showed that the patient with lower texture consistency and minimum grey-level value, which reflected a more pronounced degree of demyelination, were more likely to relapse. Therefore, the proposed PET/MR multi-modality ARR classification pipeline is a promising tool to improve the efficiency and accuracy of Computer-Aided Diagnosis (CAD) for MS.
The limitation of our study is as follows. Firstly, the dataset size in our research is relatively small due to the high cost of PET examinations and the long follow-up period required in the study. The generalization of our radiomics model needs to be tested further with a larger dataset. Secondly, the disparity between the ARR of the samples is not large enough because of the limited follow-up time. In addition, Neurofilament Light (NfL) Chains are neuron-specific intermediate proteins, which are also important biomarkers of long-term disease progression (
). A comparison of NfL and the radiomics model for ARR classification has not yet been performed in our study and will be carried out in our future study.
5. Conclusions
We proposed a workflow for predicting the ARR of patients with MS. Our automatic segmentation masks can replace manual ones with excellent performance. Furthermore, the deep learning and PET/MR radiomics-based model in our study is an effective tool in assisting ARR classification of MS patients.
Authors’ contributions
Design of the work, QZ, MZ; Data acquisition, QZ, MZ; The creation of new software used in the work, SD and CY; Data analysis, XH, HM and MC; Data investigation, HW, QH, DQ; BL and SC. Work draft, SD; Work revise, MZ, CY and SX; All authors had full access to all of the data in the study and take responsibility for the accuracy of the data and the integrity of the data analysis
Funding
This research was supported by Shanghai Pujiang Program (18PJD030), Shanghai Municipal Key Clinical Specialty (shslczdzk03403), Shanghai Shuguang Plan Project (18SG15), Shanghai outstanding young scholars Project, Shanghai talent development project (2019044), National Science Foundation of China (Grant/Award Number: 81974276, 82271383; Chinese). The funding sources had no role during the study design; in the collection, analysis, or interpretation of data; in the writing of the report; or in the decision to submit the paper for publication.
Systematic literature review and network meta-analysis in highly active relapsing-remitting multiple sclerosis and rapidly evolving severe multiple sclerosis.
Time-patterns of annualized relapse rates in randomized placebo-controlled clinical trials in relapsing multiple sclerosis: a systematic review and meta-analysis.
00254-7&id=gr1.jpg){kind=link}
00254-7&id=gr2.jpg){kind=link}
00254-7&id=gr3.jpg){kind=link}
00254-7&id=gr4.jpg){kind=link}
00254-7&id=gr5.jpg){kind=link}