1. Introduction
Multiple sclerosis (MS) is the most common demyelinating disease in humans and it affects both white matter and gray matter of the central nervous system (CNS; brain and spinal cord). The etiology of MS is not completely understood [ 1 , 2 ]. Magnetic resonance imaging (MRI) is the most common radiologic modality for noninvasive visualization of MS pathology in the CNS. MRI plays a crucial role in both diagnosis and patient management. A hallmark of MS is the presence of hyperintense lesions on T2-weighted (T2W), including T2-weighted Fluid Attenuation by Inversion Recovery (FLAIR), and proton-density-weighted (PDW) MRI. However, not all hyperintense lesions seen on T2W MRI are active. Identification of active lesions is essential for patient management [ 3 , 4 , 5 , 6 , 7 ]. Regionally compromised blood–brain barrier (BBB) is believed to be a pathologic feature of active lesions. The compromised BBB is the result of repeated venular inflammation [ 8 ]. The compromised BBB allows the leakage of gadolinium (Gd)-based contrast agents (GBCAs) from the vasculature into the brain parenchyma, which results in reduced T1 relaxation time because of the paramagnetic nature of GBCA. Consequently, on MRI, active lesions appear hyperintense on post-contrast T1-weighted (T1W) images (contrast enhancement). An association between Gd enhancement and clinical activity in MS patients has been reported [ 8 , 9 ]. Moreover, the number of volumes of enhancing lesions may be important in the evaluation of treatment efficacy [ 9 , 10 ].
Recent studies have reported that GBCA administration results in long-term accumulation of gadolinium (Gd) in various tissues [ 11 ]. While there is no definite evidence demonstrating any physiologic effect of accumulated Gd, the Food and Drug Administration issued a cautionary note that “clinicians should limit GBCA use to circumstances in which additional information provided by the contrast agent is necessary, and assess the necessity of repetitive MRIs with GBCAs” [ 12 ]. In response to these concerns, a number of techniques for identifying active lesions on unenhanced scans have been proposed [ 13 ]. Perhaps the most widely accepted method for detecting active lesions is through subtraction MRI. These methods are mainly based on the subtraction of MR images acquired at two different time points to detect expanding lesions that are considered to represent lesion activity [ 14 ]. However, all these techniques have been introduced to identify active lesions, but not quantitatively compare with the ground truth, post-contrast T1-weighted images. More recently deep learning (DL) has started making inroads in automatically identifying active lesions on unenhanced scans [ 15 ].
A number of automatic and semi-automatic techniques for identifying and segmenting enhancing lesions in MS have been reported [ 5 , 10 , 16 , 17 , 18 ]. As described by Coronado et al. [ 17 ], these methods have limitations that include the need for specialized MRI pulse sequence, extensive pre-processing, minimizing false lesion classification etc. Therefore, more recently, deep learning was used to delineate enhancing lesions in MS [ 16 , 17 , 19 ]. These DL methods are mainly based on convolutional neural networks (CNNs).
The majority of MS segmentation methods, including DL, are based on multiparametric MRI. While multiparametric-based segmentation is shown to provide excellent results, it also suffers from a few weaknesses. These include the following: (1) high dimensional feature space that increases the computational complexity, (2) ineffectiveness in explicitly capturing the spatial relationship between voxels, (3) inability to effectively capture nonlinear relationships between different voxels, since it linearly combines voxel intensities of images with different contrasts, (4) noise within MRI can compromise segmentation quality, and (5) requiring registration of multiparametric images. Many of these problems can be potentially alleviated using nonlinear dimensionality reduction (NLDR) techniques.
Machine learning is broadly divided into supervised and unsupervised learning methods [ 20 ]. Supervised learning requires labeled data created by experts to establish the ground truth. The labeling by experts could be expensive in time and cost. While CNNs are excellent at extracting global features and detecting large lesions, they may have limitations in extracting multi-scale local features, thus making it hard to detect active lesions, which may vary in size. Data-intensive tasks, like segmentation, often suffer from the curse of dimensionality, which suggests that the required sparsity and sample size needed to achieve optimal results increase with the dimensionality of the data [ 21 ]. To overcome these limitations, in this preliminary study, we have applied an unsupervised NLDR method for segmenting active lesions in MS.
Non-linear dimensionality reduction is an unsupervised machine learning technique [ 22 , 23 , 24 ]. Unlike linear dimensionality reduction (LDR), NLDR captures complex nonlinear data structures that are prominent in brain MR images. In a high-dimensional feature space, NLDR uses local features from nearest-neighbor graphs to create a lower-dimensional subspace. NLDR requires far fewer images to train than DL-based methods. This work investigates how NLDR techniques can be used to identify active MS lesions.
The objective of this study was to demonstrate that unsupervised NLDR methods outperform supervised machine learning methods in identifying active lesions. To accomplish this, we introduce key differences between LDR and NLDR. Subsequently, we establish the motivation behind applying NLDR to brain MRI data. Then, we provide details for the MR imaging dataset and performance evaluation metrics used in this work. Furthermore, we compare NLDR methods to current state-of-the-art methods for active MS lesion identification. Lastly, we discuss the potential limitations of this work.
2. Materials and Methods
2.1. Dimensionality Reduction
For combining multiparametric MRI data while maintaining inherent data structure, also called the manifold, we used dimensionality reduction (DR). A manifold is a subspace that allows the visualization of curves and surfaces through multiple coordinate systems, or charts [ 23 ]. Intuitively, a manifold is any object that can be “charted”. DR is achieved by either keeping a small subset of the most relevant information and features or by finding a smaller subset of data points [ 24 ]. DR maps high dimensional dataset X , with D dimensions, to a lower dimensional dataset Y , with d dimensions, such that d < D and X can be represented by d points.
Dimensionality reduction can be divided into two categories, LDR and NLDR, which make assumptions about X lying in a linear and nonlinear subspace, respectively, [ 25 , 26 ]. While both methods can find the lowest number of data points that are able to represent the structure of X , there are key differences between LDR and NLDR that are briefly described below.
2.2. Linear Dimensionality Reduction
LDR might reduce the complexity of high-dimensional data through a linear mapping of all data points. With LDR, we assume that the subspace that best fits X is linear. LDR has been used in several healthcare applications because it is highly adaptable and effective in extracting features from large high dimensional datasets [ 27 , 28 , 29 ]. An example of LDR is multidimensional scaling (MDS).
Multidimensional Scaling
MDS finds a linear subspace, Y, that best represents the high dimensional dataset, X, by minimizing the cost function below (1) ∑ i j ( | | x i − x i j | | − | | y i − y i j | | ) 2 such that, given the i , j -th components of X and Y , | | x i − x i j | | and | | y i − y i j | | are the respective Euclidean distances. To minimize this cost function, MDS uses eigenvalue decomposition defined as (2) B = X T ∗ X , (3) M = [ V 1 , V 2 , … V d ] , (4) Y = M ∗ X , where V are the eigenvectors, M is a feature matrix containing the first d largest eigenvalues, and B is the pairwise distances matrix of X .
Recently, Knezek et al. [ 27 ] used DR to assess various learning approaches to improve motivation amongst young learners. These authors applied MDS to add meaning and understanding to educational game play, and its effect on children’s learning performance. This work goes beyond DR for conducting statistical analysis. Malone et al. [ 28 ] proposed a Clinical Sustainability Assessment Tool (CSAT) for grouping health concepts, collected from 64 healthcare and research professionals, into several domains (e.g., engaged staff, planning, and monitoring) to create a meaningful tool for improving sustainable healthcare service delivery. Both studies used MDS to remove redundant information from high-dimensional data and create helpful insights. Although MDS effectively reduces high dimensional data, it struggles to uncover the true manifold of non-linear data, thus motivating the use of NLDR.
2.3. Non-Linear Dimensionality Reduction
We used NLDR, to uncover the true manifold of multiparametric brain MR images by reconstructing a single embedded image. While LDR assumes that the data lie in a linear subspace, NLDR can perform more intricate embeddings of nonlinear data [ 22 ]. Brain MRI is one of the primary methods for diagnosing MS, often requiring multiple image sequences to identify and characterize pathology. Brain MRI sequences, such as T2-weighted (T2W) and T1W images, provide different contrast information that aids patient management in MS. Since brain MR images share a nonlinear subspace, we employed NLDR for segmentation of active lesions. This is accomplished by reducing the dimensionality of MRI sequences into lower dimensions while preserving contrast information. Specifically, we used two NLDR algorithms, isometric feature mapping (Isomap) [ 22 , 30 ] and locally linear embedding (LLE) [ 31 ] to segment MS lesions using multiparametric MRI. The following sections will briefly describe how Isomap and LLE create a low-dimensional representation of high-dimensional data.
2.3.1. Isometric Feature Mapping
Isomap maps a high-dimensional dataset, X , to a lower-dimension dataset, Y , while retaining maximum geometric information. To reconstruct Y , Isomap first uses a K nearest neighbor graph G [ 32 ] to calculate geodesic distances (GDs) [ 33 , 34 ] of all data points of X . This creates a distance matrix, such that every data point, x i , is connected to its K nearest neighbors, x i j . Isomap then utilizes MDS to map all data points in X to a lower dimension [ 22 , 35 ]. To preserve the distance of X , MDS employs eigenvalue decomposition of the pairwise distance matrix. Specifically, we used the K neighbors classifier for the implementation of G and MDS. We used both G and MDS to reconstruct multiparametric brain MRI into a single embedded image. To maintain the manifold of X , MDS minimizes the error between pair-wise distances defined below (5) ∑ ( | | x i − x i j | | − | | y i − y i j | | ) 2 where | | x i − x i j | | and | | y i − y i j | | are the geodesic distances between data points in higher and lower dimensions, respectively. Isomap finally uses MDS mappings to create Y , a single image embedding X . The steps to create Isomap embeddings are shown in Figure 1 .
We used two shortest path algorithms to construct pairwise distance matrices, crucial for Isomap geodesic distance calculations. In Tenenbaum et al. [ 22 ], Isomap was introduced using shortest path algorithm, a recursive method which attempts to create M while holding the following recurrence relation [ 36 , 37 ], to create a geodesic distance matrix. Floyd Warshall is a recursive method which attempts to create M while holding the following recurrence relation (6) d i j k = w i j if , k = 0 min ( d i j k − 1 , d i k k − 1 , d k j k − 1 ) , if , k > 0 where w i j and d i j k represent the initialized and shortest distance lengths between points i and j , respectively, for all of data points X . Dijkstra’s is another popular shortest distance algorithm [ 37 ], which first selects a starting data point, labeled as the source, and finds the shortest path from the source to all other data points of X . According to Wang [ 37 ], Dijkstra is defined by the following relation (7) D i j = min ( D i j , D i j + d i s t ( x j , x k ) where D i j is the current distance from data point i to j , d i s t ( x j , x k ) is the distance between two abutting data points, and matrix M is created by recursively finding the pairwise distances of all data points of X . Due to the simplicity of the Floyd Warshall algorithm, it is often preferred over Dijkstra’s algorithm. Set to default in the scikit learning library, the best algorithm was automatically selected at run time, to assume optimal performance.
2.3.2. Locally Linear Embedding (LLE)
Similar to Isomap, LLE creates a mapping from X to a lower dimension while preserving its local properties. LLE assumes that embeddings can be created through small regions of X , such that any data points, x i , and its neighbors, x i j , lie on or near a linear patch of X . By assuming that the manifold is approximately linear, LLE reconstructs the properties of the data by calculating the sum of K nearest neighbors for each data point. We used G in order to calculate the Euclidean distances for all data points in X . LLE identifies coordinates of Y or y i that best align with the embedding of X [ 31 ].
To create an embedding of X , LLE characterizes the local geometry of all patches of X , as linear coefficients, and reconstructs each data point from its neighbors. The reconstruction error is defined as (8) ϵ ( W ) = ∑ n = 1 n | | x i − ∑ j = 1 k w i j x i j | | 2 , which is constrained by w i j = 0 when x i has no neighbors (i.e., x i j = 0 ) and ∑ j = 1 k w i j = 1 . To perform embedding of X in Y , LLE chooses the best coordinates of Y to minimize the embedding error defined as (9) ϕ ( Y ) = ∑ n = 1 n | | y i − ∑ j = 1 k w i j y j | | 2 . The mapping from X to Y in LLE is carried out by minimizing the reconstruction and embedding error in Equations ( 2 ) and ( 5 ).
2.3.3. Applications of Non-Linear Dimensionality Reduction
Recently in healthcare, NLDR techniques have been incorporated to extract important features from medical data. Sharma et al. [ 38 ] introduced a framework for thyroid abnormality detection, applying various dimensionality reduction techniques to ultrasound and histopathological datasets. Results based on NLDR methods outperformed current state-of-the-art computer-assisted diagnostic systems. In this work, LLE and Isomap exceeded other DR techniques, such as principal component analysis and singular value decomposition, in extracting features from ultrasound data, demonstrating that NLDR is able to reconstruct nonlinear ultrasound imaging data. We chose to use NLDR methods, LLE and Isomap, to segment active MS lesions. NLDR has been shown to outperform LDR methods in extracting features from imaging datasets [ 21 , 25 , 38 ].
In Figure 2 , we show how NLDR compares to the LDR method, MDS. Dimensions of the Swiss roll were successfully reduced to a lower dimension by using LLE and Isomap, meaning both pairwise distances of all data points and the geometric shape of the Swiss roll were preserved in the embedded images. While successfully creating a lower dimensional representation, MDS failed to unfold the manifold while maintaining its underlying structure. This suggests that LLE and Isomap effectively understand complex manifolds and can be applied to real-world data. We further validated this hypothesis on multiparametric brain MRI data and showed how LLE and Isomap are able to segment active MS lesions.
2.4. Patients
The MRI data were acquired as a part of CombiRx, a multi-center, phase 3, double-blinded, randomized clinical trial (Clinical Trial Identifier: NCT00211887 ) [ 39 ] that was supported by the National Institute of Health (NIH). Table 1 summarizes the demographics and clinical data of the clinical trial cohort at baseline. The main objective of the CombiRx clinical trial is to evaluate if treating MS patients with a combination interferon- β 1a (IFNB) and glatiramer acetate (GA) results in a better patient outcome compared to treating them with these two drugs separately. Only patients with the relapsing–remitting (RRMS) phenotype are included in this trial. Our dataset included MRIs from 46 randomly selected patients. All patients provided signed written consent and this study was approved by the local IRB. All patients had MRI-detectable MS brain lesions and were administered GBCA during the MRI scan. In total, 6 subjects were excluded due to image misregistration issues, bringing the total number of subjects to 40.
2.5. Image Dataset
Anonymized brain MR images from 46 subjects with relapsing–remitting MS were used in this work. MRIs were acquired on Philips (Best, The Netherlands) or General Electric (Milwaukee, USA), or Siemens (Erlangen, Germany) scanners operating at either 1.5T or 3T field strengths. The images included proton density (PDW), T2W, 2D FLAIR, and pre- and post-contrast T1W with voxel dimensions of (0.94 mm × 0.94 mm × 1.5 mm). All images were reviewed by two neuro MRI experts (more than 30+ years of experience) [ 40 , 41 ]. The MRI protocol and the acquisition details were provided elsewhere. Image preprocessing methods including co-registration, bias field correction and intensity normalization were applied [ 39 ].
The target lesions of this study are active white matter lesions. Grey matter lesions are not included because they rarely show enhancement. Also, grey matter lesion detection requires specialized MRI sequences such as double inversion recovery (DIR) sequence [ 42 ] and the CombiRx MRI did not include DIR imaging.
2.6. Multiparametric Brain MRI
We used LLE and Isomap to differentiate different tissue types and identify active lesions from multiparametric brain MRI. Specifically, we used PDW, FLAIR, T2W, and pre-contrast T1W as our high-dimensional dataset, and applied NLDR methods in identifying active MS lesions in brain MRI. Post-contrast T1W images were excluded from the input but were used for validation of the embedded image-based active lesion identification. The ground truth used was subtracted pre- and post-contrast T1-weighted image, referred to as the subimage. Lesions identified in subimages are labeled as active. To obtain lesion regions, we performed binary thresholding on all the ground truth and embedded images. Lower dimensional images, created by LLE and Isomap embeddings, were verified to include MS lesions by one of the authors (P.A.N).
As summarized in Figure 3 , active lesion segmentation was accomplished by (1) embedding of brain MRI images through LLE and Isomap applications, (2) calculating of the subimage, and (3) generating binary images followed by thresholding to isolate lesions. The performance of LLE and Isomap was evaluated using thresholded binary images. To quantitatively evaluate the performance of LLE and Isomap, we used Dice similarity to measure the overlap between lesion boundaries in embedded and ground truth lesions. By converting LLE and Isomap embedding to binary images, we assigned “1” to pixels related to active MS lesions, and a zero to all other pixels.
2.7. Performance Evaluation
To describe the overlap between the embedded and ground truth binary images, we use DS (10) D S = 2 A ∩ B A + B where A is the ground truth and B represents lesions identified on the binary thresholded images. (11) D S = 2 T P 2 T P + F N + F P
Equation ( 11 ), as shown in Figure 4 , TP, FP, FN, represent the true positive, false positive, and false negative sets of pixels [ 43 ], based on the overlap between A and B. Thus, if no overlap is found, DS = 0, or 1 if A and B perfectly overlap. Brain MR image resolution and the nearest neighbor value, K, were the input parameters to the model. We empirically evaluated the performance of LLE and Isomap for optimizing different input parameters.
3. Results
In order to choose the optimal input parameters for Isomap and LLE, we evaluated their sensitivity to various input resolutions and nearest neighbor values, K.
Due to high computational times in testing Isomap and LLE on original brain MR images, we initially reduced the input resolutions. However, Figure 5 shows that reducing the input resolution of LLE and Isomap degrades both methods’ ability to provide accurate DS scores due to the blurring of active lesion boundaries. Our results suggest that Isomap and LLE are able to segment active lesions given 128 2 input images but achieve optimal performance on the original 256 2 input resolution. As a result, we chose 256 2 as the input resolution for LLE and Isomap to obtain accurate DS measurements.
As shown in Figure 6 , both Isomap and LLE are insensitive to K ≥ 100. However, LLE appears to have a minor fluctuation of dice scores across k-values, and thus is more sensitive to k-values compared to Isomap, which is more stable.
To evaluate Isomap and LLE on all MS patients (n = 40), we chose K to be 100. With the chosen input resolution and K value, Isomap and LLE successfully segmented active lesions from brain MRI. Figure 7 shows single embedded images reconstructed by Isomap and LLE, while Figure 8 shows the segmentation performance of Isomap and LLE on all subjects. Isomap achieved a median DS of 0.78 ± 0.09, compared to LLE with a median DS of 0.74 ± 0.1. These results need to be further validated on a larger sample size.
4. Discussion
To the best of our knowledge, this is one of the first implementations of Isomap and LLE for gadolinium (Gd) enhancing (active) lesion segmentation. We utilized NLDR techniques to combine brain MR images into an embedded image and identify lesions. Our results show, based on DS, that both Isomap and LLE were excellent at identifying active lesions, with Isomap providing slightly better results relative to LLE. Based on our results, LLE and Isomap could potentially reduce the need for the administration of contrast agents and outperform previous state-of-the-art active lesion segmentation DL methods.
Administration of GBCA [ 7 , 44 , 45 ] is still used for identifying active lesions in MS. With recent observations about GBCA retention in various tissues, different methods have been proposed to identify Gd-enhancing lesions, either by reducing the amount of GBCA administered or eliminating it. However, the majority of published methods, particularly those based on DL, utilized multiparametric MRI and supervised learning. These methods have their own disadvantages. In this preliminary study, we attempted to overcome these disadvantages by using unsupervised techniques, Isomap, and LLE. Our results suggest that NLDR methods if confirmed on a larger sample size, can be deployed in a clinical setting.
Multiple MRI studies are reported on the detection of active lesions on unenhanced scans. The results of subtraction images are compared with post-T1W images to assess the accuracy. However, they did not perform lesion segmentation [ 7 , 14 , 46 , 47 , 48 ]. Thus, it is very difficult to compare our results with these published ones. On the other hand, there are existing AI-based clinical applications for active lesion identification. In [ 7 ], Rudie et al. identified active MS lesions on non-contrast images using a computer-assisted-detection (CAD)-based system. This work demonstrated that it is possible to perform real-time assessments of non-contrast images to determine the necessity of contrast administration. However, this work requires manual annotations of MS lesions to achieve optimal performance. LLE and Isomap achieved excellent performance in identifying active MS lesions without need for manual annotated trained images as input. Identifying active lesions without contrast agents and less supervision could potentially improve patient safety and reduce the cost of clinical care.
Based on Coronado et al. [ 17 ], our work appears to outperform current state-of-the-art (SOTA) DL methods in segmenting active lesions in brain MRI. However, this work included GBCA imaging while we excluded post-contrast T1W images from testing. Current methods for active MS lesion segmentation use deep convolutional neural networks (CNNs) [ 17 , 19 , 40 , 41 ] to segment active lesions from brain MRI. Coronado et al. achieved a mean dice similarity score of 0.77 which is considered to be excellent. Our method offered a very comparable score of 0.78. Our method offers a number of advantages over the CNN method, which typically needs a large amount of annotated imaging data by experts, which could be very time-consuming and not feasible for large data. Although this work achieved dice similarity performance as shown in this preliminary study, our method is an unsupervised machine learning technique and does not require using a subset of data for training and the entire dataset could be used as a validation set.
Instead of using manual lesion annotations to guide reconstruction, LLE and Isomap use local image features from K nearest neighbor graphs to reconstruct high dimensional inputs in a lower dimensional space. This suggests NLDR performance is dependent on local topological features [ 22 , 32 ]. Other segmentation studies of breast [ 25 ], and ischemic lesions in brain MRI [ 49 ], used Isomap and LLE to reconstruct multiparametric MRI data with fewer than 30 subjects.
Akhbardeh et al. [ 25 ] used DR to embed multiple breast MR images and segment breast lesions. The performance of linear and nonlinear DR methods was compared to determine the best method for breast tissue segmentation. Both LLE and Isomap outperformed LDR methods, principal component analysis and MDS, and resulted in excellent similarity to post-contrast T1-weighted images that are consistent with our results. In this work, we exploited the robustness of NLDR methods for segmenting smaller embedded MS lesions compared to larger breast lesions.
To segment tissue at risk in stroke patients and quantify the amount of salvageable brain tissue, Parekh et al. [ 49 ] integrated multiparametric stroke MRI data using LLE and Isomap, demonstrating high similarity to radiological gold standards. However, this study was limited to animal MRI data, and thus needs to be further validated on human subjects. With the application of NLDR in early cerebral ischemia detection, this work shows NLDR methods as potential solutions for other diseases that require multiparametric MRI data to perform accurate diagnosis. Similarly, with brain MRI data, Park [ 50 ] used Isomap to quantify the shape of brain MRI and detection of Alzheimer’s Disease. This work revealed that Isomap can delineate the true shape of brain MR images, distinguish between AD and normal manifolds, and potentially be used for brain MRI time-series analysis.
There are potential limitations in our study. For example, there is increased computational complexity, in both Isomap and LLE, when reconstructing embedded images. This is largely due to shortest path distance and eigenvalue calculations. This limitation could be overcome using a subset, “landmarks”, of high dimensional data points to reduce the computations required for both Isomap and LLE [ 51 , 52 , 53 , 54 ].
Another limitation is that our sample size was 40 subjects. Further validation is needed on a larger and independent sample size that is drawn from a different distribution. Lastly, this work does not measure the sensitivity of LLE and Isomap to different lesion sizes. Investigating and addressing active MS lesion identification performance as a function of lesion sizes could further improve the performance of LLE and Isomap.
5. Conclusions
In conclusion, we demonstrated unsupervised learning techniques, LLE and Isomap, are able to capture the manifold of multiparametric MRI data and segment active MS lesions. Although both methods exhibited excellent performance, Isomap slightly outperformed LLE and showed more consistent results in the sensitivity analysis of input parameters. LLE and Isomap potentially could assist in the clinical management of MS patients, and reduce the usage of GBCAs for identifying active lesions in brain MRIs.
6. Patents
A.A. and M.A.J. have patent, “Multiparametric non-linear dimension reduction methods and systems related thereto”, US Patent 9,256,966.
Acknowledgments
We thank Jerry S. Wolinsky and Fred D. Lublin for providing access to the CombiRx MRI data.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Author Contributions
Methodology, A.A. and M.A.J.; software and code, A.A. and J.U.; validation, A.A., M.A.J. and J.U.; formal analysis, J.U. and A.A.; resources, M.A.J., A.A., P.A.N. and V.B.; data curation, P.A.N. and A.A.; writing—original draft preparation, J.U. and A.A.; writing—review and editing, J.U., M.A.J., P.A.N., A.K., A.A. and V.B.; visualization, J.U.; supervision, A.A., M.A.J., P.A.N. and V.B.; project administration, A.A., M.A.J. and V.B.; funding acquisition, V.B., A.A. and M.A.J. served as the driving force and guidance in implementing the project, data analysis, and results throughout the execution of the project. All authors have read and agreed to the published version of the manuscript.
Institutional Review Board Statement
This research was approved by the Committee for the Protection of Human Subjects of the University of Texas Health Science Center Houston and started on January 2005 and completed in March 2013. It was a double-blind, multicenter phase III clinical trial ( clinicaltrials.gov code is NCT00211887 ).
Informed Consent Statement
Patient consent was waived due to the fact that no patient personal information was retrieved or processed off-site.
Data Availability Statement
Data will be made available upon request, and availability is determined by the institutional guidelines.
Conflicts of Interest
The authors declare no conflicts of interest.
References
1. Dobson R. Giovannoni G. Multiple sclerosis—A review Eur. J. Neurol. 2019 26 27 40 10.1111/ene.13819 30300457
2. Dilokthornsakul P. Valuck R.J. Nair K.V. Corboy J.R. Allen R.R. Campbell J.D. Multiple sclerosis prevalence in the United States commercially insured population Neurology 2016 86 1014 1021 10.1212/WNL.0000000000002469 26888980 PMC4799713
3. Sadigh G. Saindane A. Waldman A. Lava N. Hu R. Comparison of unenhanced and gadolinium-enhanced imaging in multiple sclerosis: Is contrast needed for routine follow-up MRI? Am. J. Neuroradiol. 2019 40 1476 1480 10.3174/ajnr.A6179 31439627 PMC7048442
4. Saade C. Bou-Fakhredin R. Yousem D.M. Asmar K. Naffaa L. El-Merhi F. Gadolinium and multiple sclerosis: Vessels, barriers of the brain, and glymphatics Am. J. Neuroradiol. 2018 39 2168 2176 10.3174/ajnr.A5773 30385472 PMC7655403
5. Datta S. Sajja B.R. He R. Gupta R.K. Wolinsky J.S. Narayana P.A. Segmentation of gadolinium-enhanced lesions on MRI in multiple sclerosis J. Magn. Reson. Imaging Off. J. Int. Soc. Magn. Reson. Med. 2007 25 932 937 10.1002/jmri.20896 17457804
6. Warntjes J. Tisell A. Landtblom A.M. Lundberg P. Effects of gadolinium contrast agent administration on automatic brain tissue classification of patients with multiple sclerosis Am. J. Neuroradiol. 2014 35 1330 1336 10.3174/ajnr.A3890 24699093 PMC7966571
7. Rudie J.D. Mattay R.R. Schindler M. Steingall S. Cook T.S. Loevner L.A. Schnall M.D. Mamourian A.C. Bilello M. An Initiative to Reduce Unnecessary Gadolinium-Based Contrast in Multiple Sclerosis Patients J. Am. Coll. Radiol. 2019 16 1158 1164 10.1016/j.jacr.2019.04.005 31092348 PMC6732018
8. Karimian-Jazi K. Wildemann B. Diem R. Schwarz D. Hielscher T. Wick W. Bendszus M. Breckwoldt M.O. Gd contrast administration is dispensable in patients with MS without new T2 lesions on follow-up MRI Neurol.-Neuroimmunol. Neuroinflamm. 2018 5 e480 10.1212/NXI.0000000000000480 30038948 PMC6053940
9. Rovira À. Doniselli F.M. Auger C. Haider L. Hodel J. Severino M. Wattjes M.P. van der Molen A.J. Jasperse B. Mallio C.A. Use of gadolinium-based contrast agents in multiple sclerosis: A review by the ESMRMB-GREC and ESNR Multiple Sclerosis Working Group Eur. Radiol. 2023 34 1726 1735 37658891 10.1007/s00330-023-10151-y
10. Zeng C. Gu L. Liu Z. Zhao S. Review of deep learning approaches for the segmentation of multiple sclerosis lesions on brain MRI Front. Neuroinform. 2020 14 610967 10.3389/fninf.2020.610967 33328949 PMC7714963
11. Starekova J. Pirasteh A. Reeder S.B. Update on Gadolinium Based Contrast Agent Safety, From the AJR Special Series on Contrast Media Am. J. Roentgenol. 2023 10.2214/AJR.23.30036 37850581
12. FDA Drug Safety Communication FDA Warns That Gadolinium-Based Contrast Agents (GBCAs) Are Retained in the Body 2018 Available online: https://www.fda.gov/drugs/drug-safety-and-availability/fda-drug-safety-communication-fda-warns-gadolinium-based-contrast-agents-gbcas-are-retained-body (accessed on 1 March 2024)
13. Eichinger P. Schön S. Pongratz V. Wiestler H. Zhang H. Bussas M. Hoshi M.M. Kirschke J. Berthele A. Zimmer C. Accuracy of unenhanced MRI in the detection of new brain lesions in multiple sclerosis Radiology 2019 291 429 435 10.1148/radiol.2019181568 30860448
14. Moraal B. Meier D.S. Poppe P.A. Geurts J.J. Vrenken H. Jonker W.M. Knol D.L. van Schijndel R.A. Pouwels P.J. Pohl C. Subtraction MR images in a multiple sclerosis multicenter clinical trial setting Radiology 2009 250 506 514 10.1148/radiol.2501080480 19037018 PMC2657481
15. Narayana P.A. Coronado I. Sujit S.J. Wolinsky J.S. Lublin F.D. Gabr R.E. Deep Learning for Predicting Enhancing Lesions in Multiple Sclerosis from Noncontrast MRI Radiology 2020 294 398 404 10.1148/radiol.2019191061 31845845 PMC6980901
16. Schlaeger S. Shit S. Eichinger P. Hamann M. Opfer R. Krüger J. Dieckmeyer M. Schön S. Mühlau M. Zimmer C. AI-based detection of contrast-enhancing MRI lesions in patients with multiple sclerosis Insights Into Imaging 2023 14 123 10.1186/s13244-023-01460-3 37454342 PMC10350445
17. Coronado I. Gabr R.E. Narayana P.A. Deep learning segmentation of gadolinium-enhancing lesions in multiple sclerosis Mult. Scler. J. 2021 27 519 527 10.1177/1352458520921364 PMC7680286 32442043
18. Karimaghaloo Z. Shah M. Francis S.J. Arnold D.L. Collins D.L. Arbel T. Automatic detection of gadolinium-enhancing multiple sclerosis lesions in brain MRI using conditional random fields IEEE Trans. Med. Imaging 2012 31 1181 1194 10.1109/TMI.2012.2186639 22318484
19. Gaj S. Ontaneda D. Nakamura K. Automatic segmentation of gadolinium-enhancing lesions in multiple sclerosis using deep learning from clinical MRI PLoS ONE 2021 16 e0255939 10.1371/journal.pone.0255939 34469432 PMC8409666
20. Russell S.J. Norvig P. Artificial Intelligence: A Modern Approach 3rd ed. Pearson London, UK 2010
21. Liu N. Chee M.L. Koh Z.X. Leow S.L. Ho A.F.W. Guo D. Ong M.E.H. Utilizing machine learning dimensionality reduction for risk stratification of chest pain patients in the emergency department BMC Med. Res. Methodol. 2021 21 74 10.1186/s12874-021-01265-2 33865317 PMC8052947
22. Tenenbaum J.B. Silva V.d. Langford J.C. A global geometric framework for nonlinear dimensionality reduction Science 2000 290 2319 2323 10.1126/science.290.5500.2319 11125149
23. Sorzano C.O.S. Vargas J. Montano A.P. A survey of dimensionality reduction techniques arXiv 2014 1403.2877
24. Meilă M. Zhang H. Manifold learning: What, how, and why Annu. Rev. Stat. Its Appl. 2023 11 10.1146/annurev-statistics-040522-115238
25. Akhbardeh A. Jacobs M.A. Comparative analysis of nonlinear dimensionality reduction techniques for breast MRI segmentation Med. Phys. 2012 39 2275 2289 10.1118/1.3682173 22482648 PMC3337666
26. Linting M. Meulman J.J. Groenen P.J. van der Koojj A.J. Nonlinear principal components analysis: Introduction and application Psychol. Methods 2007 12 336 10.1037/1082-989X.12.3.336 17784798
27. Knezek G. Gibson D. Christensen R. Trevisan O. Carter M. Assessing approaches to learning with nonparametric multidimensional scaling Br. J. Educ. Technol. 2023 54 126 141 10.1111/bjet.13275 37035265 PMC10078274
28. Malone S. Prewitt K. Hackett R. Lin J.C. McKay V. Walsh-Bailey C. Luke D.A. The clinical sustainability assessment tool: Measuring organizational capacity to promote sustainability in healthcare Implement. Sci. Commun. 2021 2 1 12 10.1186/s43058-021-00181-2 34274004 PMC8285819
29. Patra S.S. Harshvardhan G. Gourisaria M.K. Mohanty J.R. Choudhury S. Emerging healthcare problems in high-dimensional data and dimension reduction Advanced Prognostic Predictive Modelling in Healthcare Data Analytics Springer Singapore 2021 25 49
30. Law M.H. Jain A.K. Incremental nonlinear dimensionality reduction by manifold learning IEEE Trans. Pattern Anal. Mach. Intell. 2006 28 377 391 10.1109/TPAMI.2006.56 16526424
31. Roweis S.T. Saul L.K. Nonlinear dimensionality reduction by locally linear embedding Science 2000 290 2323 2326 10.1126/science.290.5500.2323 11125150
32. Kim J. Kim B.S. Savarese S. Comparing image classification methods: K-nearest-neighbor and support-vector-machines Proceedings of the 6th WSEAS International Conference on Computer Engineering and Applications, and Proceedings of the 2012 American Conference on Applied Mathematics Stevens Point, WI, USA 25–27 January 2012 133 138
33. Shamai G. Kimmel R. Geodesic distance descriptors Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Honolulu, HI, USA 21–26 July 2017 6410 6418
34. Crane K. Livesu M. Puppo E. Qin Y. A survey of algorithms for geodesic paths and distances arXiv 2020 2007.10430
35. Steyvers M. Multidimensional scaling Encyclopedia of Cognitive Science Nature Publishing Group Berlin, Germany 2002 Volume 1 160 165
36. Toroslu I.H. Improving the floyd-warshall all pairs shortest paths algorithm arXiv 2021 2109.01872
37. Wang X.Z. The comparison of three algorithms in shortest path issue J. Phys. Conf. Ser. 2018 1087 022011 10.1088/1742-6596/1087/2/022011
38. Sharma R. Mahanti G.K. Panda G. Rath A. Dash S. Mallik S. Hu R. A framework for detecting thyroid cancer from ultrasound and histopathological images using deep learning, meta-heuristics, and MCDM algorithms J. Imaging 2023 9 173 10.3390/jimaging9090173 37754937 PMC10532397
39. Lublin F.D. Cofield S.S. Cutter G.R. Conwit R. Narayana P.A. Nelson F. Salter A.R. Gustafson T. Wolinsky J.S. Investigators C. Randomized study combining interferon and glatiramer acetate in multiple sclerosis Ann. Neurol. 2013 73 327 340 10.1002/ana.23863 23424159 PMC3631288
40. Sajja B.R. Datta S. He R. Mehta M. Gupta R.K. Wolinsky J.S. Narayana P.A. Unified approach for multiple sclerosis lesion segmentation on brain MRI Ann. Biomed. Eng. 2006 34 142 151 10.1007/s10439-005-9009-0 16525763 PMC1463248
41. Datta S. Sajja B.R. He R. Wolinsky J.S. Gupta R.K. Narayana P.A. Segmentation and quantification of black holes in multiple sclerosis Neuroimage 2006 29 467 474 10.1016/j.neuroimage.2005.07.042 16126416 PMC1808226
42. Bedell B.J. Narayana P.A. Implementation and evaluation of a new pulse sequence for rapid acquisition of double inversion recovery images for simultaneous suppression of white matter and CSF J. Magn. Reson. Imaging 1998 8 544 547 10.1002/jmri.1880080305 9626866
43. Bokhovkin A. Burnaev E. Boundary loss for remote sensing imagery semantic segmentation Proceedings of the International Symposium on Neural Networks Moscow, Russia 10–12 July 2019 Springer Berlin, Germany 2019 388 401
44. Falk Delgado A. Van Westen D. Nilsson M. Knutsson L. Sundgren P.C. Larsson E.M. Falk Delgado A. Diagnostic value of alternative techniques to gadolinium-based contrast agents in MR neuroimaging—A comprehensive overview Insights Into Imaging 2019 10 1 15 10.1186/s13244-019-0771-1 31444580 PMC6708018
45. Rogosnitzky M. Branch S. Gadolinium-based contrast agent toxicity: A review of known and proposed mechanisms Biometals 2016 29 365 376 10.1007/s10534-016-9931-7 27053146 PMC4879157
46. Donatelli G. Cecchi P. Migaleddu G. Cencini M. Frumento P. D’Amelio C. Peretti L. Buonincontri G. Pasquali L. Tosetti M. Quantitative T1 mapping detects blood–brain barrier breakdown in apparently non-enhancing multiple sclerosis lesions NeuroImage Clin. 2023 40 103509 10.1016/j.nicl.2023.103509 37717382 PMC10514220
47. Liguori M. Meier D.S. Hildenbrand P. Healy B.C. Chitnis T. Baruch N.F. Khoury S.J. Weiner H.L. Bakshi R. Barkhof F. One year activity on subtraction MRI predicts subsequent 4 year activity and progression in multiple sclerosis J. Neurol. Neurosurg. Psychiatry 2011 82 1125 1131 10.1136/jnnp.2011.242115 21429902
48. Schlaeger S. Li H.B. Baum T. Zimmer C. Moosbauer J. Byas S. Mühlau M. Wiestler B. Finck T. Longitudinal Assessment of Multiple Sclerosis Lesion Load With Synthetic Magnetic Resonance Imaging—A Multicenter Validation Study Investig. Radiol. 2023 58 320 326 10.1097/RLI.0000000000000938 36730638
49. Parekh V.S. Jacobs J.R. Jacobs M.A. Unsupervised nonlinear dimensionality reduction machine learning methods applied to multiparametric MRI in cerebral ischemia: Preliminary results Proceedings of the Medical Imaging 2014: Image Processing San Diego, CA, USA 16–18 February 2014 Volume 9034 681 689
50. Park H. The ADNI ISOMAP induced manifold embedding and its application to Alzheimer’s disease and mild cognitive impairment Neurosci. Lett. 2012 513 141 145 10.1016/j.neulet.2012.02.016 22366258
51. Silva V. Tenenbaum J. Global versus local methods in nonlinear dimensionality reduction Adv. Neural Inf. Process. Syst. 2002 15 721 728
52. Chen Y. Crawford M. Ghosh J. Improved nonlinear manifold learning for land cover classification via intelligent landmark selection Proceedings of the 2006 IEEE International Symposium on Geoscience and Remote Sensing Denver, CO, USA 31 July–4 August 2006 545 548
53. Shi H. Yin B. Bao Y. Lei Y. A novel landmark point selection method for L-ISOMAP Proceedings of the 2016 12th IEEE International Conference on Control and Automation (ICCA) Kathmandu, Nepal 1–3 June 2016 621 625
54. Braverman V. Feldman D. Lang H. Statman A. Zhou S. Efficient coreset constructions via sensitivity sampling Proceedings of the Asian Conference on Machine Learning Virtual 17–19 November 2021 948 963