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
Differentiation between multiple sclerosis and neuromyelitis optica spectrum disorders by multiparametric quantitative MRI using convolutional neural network
Department of Radiology, Juntendo University School of Medicine, 1-2-1, Hongo, Bunkyo-ku, Tokyo 113-8421, JapanMilliman Inc. Urbannet Kojimachi Building 8F, 1-6-2 Kojimachi, Tokyo 102-0083, JapanPlusman LLC, 2F 1-3-6 Hirakawacho, Chiyoda-ku, Tokyo 102-0093, Japan
Department of Radiology, Juntendo University School of Medicine, 1-2-1, Hongo, Bunkyo-ku, Tokyo 113-8421, JapanDepartment of Radiology, Graduate School of Medicine, The University of Tokyo, 7-3-1, Hongo, Bunkyo-ku, Tokyo 113-8655, Japan
Department of Radiology, Juntendo University School of Medicine, 1-2-1, Hongo, Bunkyo-ku, Tokyo 113-8421, JapanDepartment of Radiology, Toho University Omori Medical Center, 6-11-1 Omorinishi, Ota-ku, Tokyo 143-8541, Japan
Department of Radiology, Juntendo University School of Medicine, 1-2-1, Hongo, Bunkyo-ku, Tokyo 113-8421, JapanDepartment of Radiology, Graduate School of Medicine, The University of Tokyo, 7-3-1, Hongo, Bunkyo-ku, Tokyo 113-8655, Japan
Differentiation between multiple sclerosis and neuromyelitis optica spectrum disorders is crucial.
•
A convolutional neural network was developed to differentiate between these two diseases.
•
The network is based on multiparametric quantitative MRI.
•
The area under the receiver operating characteristic curve of the model was 0.859.
Abstract
Multiple sclerosis and neuromyelitis optica spectrum disorders are both neuroinflammatory diseases and have overlapping clinical manifestations. We developed a convolutional neural network model that differentiates between the two based on magnetic resonance imaging data. Thirty-five patients with relapsing-remitting multiple sclerosis and eighteen age-, sex-, disease duration-, and Expanded Disease Status Scale-matched patients with anti-aquaporin-4 antibody-positive neuromyelitis optica spectrum disorders were included in this study. All patients were scanned on a 3-T scanner using a multi-dynamic multi-echo sequence that simultaneously measures R1 and R2 relaxation rates and proton density. R1, R2, and proton density maps were analyzed using our convolutional neural network model. To avoid overfitting on a small dataset, we aimed to separate features of images into those specific to an image and those common to the group, based on SqueezeNet. We used only common features for classification. Leave-one-out cross validation was performed to evaluate the performance of the model. The area under the receiver operating characteristic curve of the developed convolutional neural network model for differentiating between the two disorders was 0.859. The sensitivity to multiple sclerosis and neuromyelitis optica spectrum disorders, and accuracy were 80.0%, 83.3%, and 81.1%, respectively. In conclusion, we developed a convolutional neural network model that differentiates between multiple sclerosis and neuromyelitis optica spectrum disorders, and which is designed to avoid overfitting on small training datasets. Our proposed algorithm may facilitate a differential diagnosis of these diseases in clinical practice.
Multiple sclerosis (MS) and neuromyelitis optica spectrum disorders (NMOSD) are neuroinflammatory diseases with overlapping clinical manifestations and imaging features, which makes it difficult to distinguish the two [
]. It is clinically important to differentiate between these diseases because treatment options and prognoses may differ dramatically. Notably, standard MS therapy, such as interferon-beta, dimethyl fumarate, fingolimod, and natalizumab, may exacerbate NMOSD and increase relapse rates [
Comparison of myelin water fraction values in periventricular white matter lesions between multiple sclerosis and neuromyelitis optica spectrum disorder.
SyMRI of the brain: rapid quantification of relaxation rates and proton density, with synthetic MRI, automatic brain segmentation, and myelin measurement.
], has the potential to objectively distinguish between these diseases. However, an automated method that allows prompt, repeatable, and reproducible analyses of quantitative MRI data in a clinical setting has not been fully established. The convolutional neural network (CNN) has been developed for fully automated image feature extraction and classification, and has been applied to MRI for differentiation of brain diseases, such as diseases presenting with cognitive impairments [
A novel deep learning approach with a 3D convolutional ladder network for differential diagnosis of idiopathic normal pressure hydrocephalus and Alzheimer's Disease.
]. However, this method has not been utilized for differentiating between MS and NMOSD. Herein, we propose a CNN algorithm that helps to efficiently distinguish between MS and NMOSD based on the quantitative values acquired with quantitative synthetic MRI.
2. Methods
2.1 Participants and MRI
Thirty-five patients with relapsing-remitting MS and eighteen patients with anti-aquaporin-4 antibody-positive NMOSD were prospectively included in this study from December 2016 to October 2018. NMOSD cases were consecutive and MS cases were selected so that these two groups were age-, sex-, disease-duration-, and Expanded Disease Status Scale-matched (Table 1). The diagnosis of MS was made according to the 2010 revised McDonald criteria [
] on a 3-T MR scanner (Discovery 750w, GE Healthcare, Waukesha, WI, USA). Multi-dynamic multi-echo sequence was used with combinations of 2 echo times (16.9 and 84.5 ms) and 4 saturation-delay times (146, 546, 1879, and 3879 ms). The other parameters were as follows: repetition time = 4.0 s, field of view = 240 × 240 mm, matrix = 320 × 320, echo-train length = 10, bandwidth = 31.25 kHz, section thickness/gap = 4.0 mm/1.0 mm, slices = 30, and acquisition time = 7 min 12 s. To extract R1, R2, and PD maps while accounting for B1 inhomogeneity, a least square fit was performed on the signal intensity (SI) of the 8 source images by minimizing the following equation:
where α is the applied excitation flip angle 90° and θ is the saturation flip angle of 120°. A is an overall intensity scaling factor that takes into account several elements, including sensitivity of the coil, amplification of the radiofrequency chain, and voxel volume. The details of the sequence composition and postprocessing are described elsewhere [
SyMRI of the brain: rapid quantification of relaxation rates and proton density, with synthetic MRI, automatic brain segmentation, and myelin measurement.
]. The CNN training in the following section was performed on R1, R2, and PD maps masked by the intracranial volume automatically segmented on SyMRI software [
SyMRI of the brain: rapid quantification of relaxation rates and proton density, with synthetic MRI, automatic brain segmentation, and myelin measurement.
]. This study was conducted in accordance with the Declaration of Helsinki and reviewed and approved by the institutional review board (approval number 15-212). All patients enrolled completed the informed consent form.
Table 1Demographics of patients.
MS
NMOSD
P value
N
35
18
Sex (F/M)
29/6
16/2
0.56
Age (years), mean ± SD
50.7 ± 8.0
52.7 ± 16.1
0.56
Disease Duration (years), mean ± SD
11.9 ± 4.6
11.1 ± 5.8
0.60
Median EDSS [range]
2 [0–7]
3 [0–6]
0.14
P values correspond to χ2 for sex, t-test for age and disease duration, and Mann-Whitney U test for EDSS. EDSS = Expanded Disease Status Scale; MS = multiple sclerosis; NMOSD = neuromyelitis optica spectrum disorders.
Fig. 1Representative images of patients with MS and NMOSD. R1, R2, and PD maps, with corresponding synthetic FLAIR images (post-processing repetition time = 15,000 ms, echo time = 100 ms, and inversion time = 3000 ms) created from R1, R2, and PD maps are shown for MS (top row) and NMOSD (bottom row) patients. These two cases were correctly differentiated by the CNN model in this study. PD is shown as pu, where PD of pure water at 37 °C corresponds to 100 pu. CNN = convolutional neural network; FLAIR = fluid-attenuated inversion recovery; MS = multiple sclerosis; NMOSD = neuromyelitis optica spectrum disorders; PD = proton density; pu = percent unit.
Iandola FN, Moskewicz MW, Ashraf K, Han S, Dally WJ, Keutzer K. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size. arXiv:160207360. 2016.
], which can be performed with fewer parameters than typical CNN models for extracting features from the input. Additionally, SqueezeNet uses global average pooling instead of fully connected layers that are prone to overfitting. The model design is illustrated in Fig. 2. To further avoid overfitting, we separated features of images into those specific only to each image and those common to the groups (i.e. MS or NMOSD). The underlying idea is that the fitting of a model to features specific to each image causes overfitting. We assumed that channels of the last convolutional layer represent features of the image, and the outstanding channels through the stack of images were considered as features of the patient. An outstanding channel was defined as a channel whose maximum pixel in the feature maps after convolution exceeds the mean plus two standard deviations of the maximums of the whole channels of each image. Sum-set operation was performed for the outstanding channels in all 30 slices of each patient. The outstanding channels common to at least two-thirds of each group and less than one-eighth of the other group were defined as common features of the group. These ratios were empirically determined. We used only common features for classification of the patients. Since such common features’ channels change along with training epochs, the training model was constructed so that it evaluates and updates common features’ channels at each epoch. The output scores, which were vectors with 2 values due to the global average pooling, were averaged on 30 slices for each patient. The softmax of this average value was considered to be the prediction of each patient. The loss function was minimized by the Adam rule [
] with α = 0.0002, β1 = 0.9, and β2 = 0.999. The batch size was set at 1. The training was stopped when the accuracy of prediction on the training data exceeded 0.98. The program was coded using Python 3.6 with Chainer 6.4.0 (Python Software Foundation). The model performance was evaluated by leave-one-out cross-validation for all 53 cases. To demonstrate the effect of common feature selection, we also developed and evaluated a model without common feature selection.
Fig. 2The proposed CNN architecture for MS and NMOSD classification. We designed a CNN model that extracts features common to each group. We used only outstanding channels common to the groups for classification of each case. Since such common features’ channels change along with training epochs, the training model was constructed so that it evaluates and updates common features’ channels. CNN = convolutional neural network; MS = multiple sclerosis; NMOSD = neuromyelitis optica spectrum disorders.
Receiver operating characteristic (ROC) analysis was performed to assess the diagnostic performance of the CNN model. All statistical analyses were performed with GraphPad Prism (Version 8.4.2; GraphPad Software, La Jolla, CA).
5. Results
Among all 53 models used for leave-one-out validation, the mean ± standard deviation of the percentage of common features among all channels was 18% ± 14%. The area under the ROC curve of the diagnostic ability of the developed CNN, which was used to differentiate between MS and NMOSD, was 0.859 (Fig. 3), which was much higher than the value 0.572 for the model without common feature selection. When the cutoff was set at the value where the sum of the sensitivity to MS and NMO was the highest, the sensitivity to MS and NMOSD was 80.0% and 83.3%, respectively (Fig. 3). The accuracy was 81.1%.
Fig. 3Receiver operating characteristic curve of the diagnostic ability by the CNN model for differentiating between MS and NMOSD. The red circle indicates the point where the sum of the sensitivity to MS and NMOSD is the highest. CNN = convolutional neural network; MS = multiple sclerosis; NMOSD = neuromyelitis optica spectrum disorders.
In the present study, we designed a CNN model to overcome the overfitting caused by the small size of the training data. To our knowledge, this is the first study using CNN to differentiate between MS and NMOSD, and we here demonstrated that our proposed model achieved an accuracy of 81.1% in this context. Thus, it was slightly higher than that observed in a previous study, which reported an accuracy of 80% using 3D T1-weighted and 3D fluid-attenuated inversion recovery (FLAIR) images as inputs of a random forest algorithm [
]. However, the algorithm used in that study requires feature extraction as pre-processing, which is difficult and time-consuming in a clinical setting. In contrast, our deep learning model automatically performs both feature extraction and classification. Further, we used only multi-dynamic multi-echo data in this study. A previous study that combined functional MRI and diffusion tensor imaging with T1-weighted and FLAIR images for a support vector machine algorithm, achieved an accuracy of 88% in distinguishing MS and NMOSD [
], but these advanced MRI modalities are difficult to be incorporated into routine clinical scans. The multi-dynamic multi-echo sequence is now a clinical routine in some institutions, because conventional contrast-weighted images, such as T1-weighted, T2-weighted, and FLAIR images, and even magnetic resonance angiography, can be created using synthetic MRI [
SyMRI of the brain: rapid quantification of relaxation rates and proton density, with synthetic MRI, automatic brain segmentation, and myelin measurement.
]. Including other modalities that are clinically available, such as lumbar puncture results, antibody status, and quantitative spinal cord measures, may further increase the performance of our algorithm. Our algorithm may also be applicable to other quantitative imaging methods, such as magnetic resonance fingerprinting [
], once trained by datasets incorporating these disorders.
The major limitation of our study is the small size of the dataset without an external test dataset. The generalizability of our model should be confirmed using a larger dataset in a future study. Nonetheless, our model is expected to be fairly generalizable to other scanners, because high inter-scanner reproducibility of quantitative values has been reported for multi-dynamic multi-echo sequence [
Linearity, bias, intrascanner repeatability, and interscanner reproducibility of quantitative multidynamic multiecho sequence for rapid simultaneous relaxometry at 3 T: a validation study with a standardized phantom and healthy controls.
]. Another limitation is the lack of healthy controls. Hence, for the time being, the proposed algorithm should be used only for subjects with clinical findings suspicious of MS and/or NMOSD. Lastly, the ratios used for common features to extract outstanding channels were empirically determined. These can be considered as hyperparameters and may not be optimal for other data.
7. Conclusions
We developed a CNN model differentiating between MS and NMOSD designed to avoid overfitting in small training datasets. However, the extent to which our model can be transferred to other systems should be confirmed using an external dataset.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgement
This work was supported by Japan Agency for Medical Research and Development (AMED) under Grant Number JP19lk1010025h9902; JSPS KAKENHI grant number 19K17150, 19K17177, 18K07692, and 18H02772; a Grant-in-Aid for Special Research in Subsidies for ordinary expenses of private schools from The Promotion and Mutual Aid Corporation for Private Schools of Japan; Brain/MINDS Beyond program from AMED Grant Number JP19dm0307024 and JP19dm0307101.
Comparison of myelin water fraction values in periventricular white matter lesions between multiple sclerosis and neuromyelitis optica spectrum disorder.
SyMRI of the brain: rapid quantification of relaxation rates and proton density, with synthetic MRI, automatic brain segmentation, and myelin measurement.
A novel deep learning approach with a 3D convolutional ladder network for differential diagnosis of idiopathic normal pressure hydrocephalus and Alzheimer's Disease.
Iandola FN, Moskewicz MW, Ashraf K, Han S, Dally WJ, Keutzer K. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size. arXiv:160207360. 2016.
Linearity, bias, intrascanner repeatability, and interscanner reproducibility of quantitative multidynamic multiecho sequence for rapid simultaneous relaxometry at 3 T: a validation study with a standardized phantom and healthy controls.
00091-6&id=gr1.jpg){kind=link}
00091-6&id=gr2.jpg){kind=link}
00091-6&id=gr3.jpg){kind=link}