Source-linked AI summary
Applications of Deep Learning Techniques for Automated Multiple Sclerosis Detection Using Magnetic Resonance Imaging: A Review
Afshin Shoeibi, Marjane Khodatars, Mahboobeh Jafari, Parisa Moridian, Mitra Rezaei, Roohallah Alizadehsani, Fahime Khozeimeh, Juan Manuel Gorriz, Jónathan Heras, Maryam Panahiazar, Saeid Nahavandi, U. Rajendra Acharya
TL;DR
MS diagnosis from MRI is clinically valuable but time-consuming and difficult to perform manually. This review synthesizes deep-learning methods and MRI datasets for automated MS detection, then examines research patterns, challenges, and future directions. The reviewed literature emphasizes segmentation and classification, with CNNs widely used; key constraints include limited datasets, unavailable functional MRI data, and hardware or privacy barriers.
Problem
MRI-based MS diagnosis is time-consuming and challenging, while manual segmentation and classification of lesions can be problematic and prone to misdiagnosis.
Method
The paper reviews DL-based CAD systems, MRI datasets, architectures, preprocessing methods, and applications for automated MS lesion segmentation and MS detection.
Results
The reviewed studies predominantly use segmentation and classification, with CNN models widely employed for MS diagnosis from MRI modalities.
Takeaways & Limitations
DL methods provide a framework for automated MS lesion localization and detection using MRI, supporting clinicians’ diagnostic assessment within the reviewed scope.
Takeaways & Limitations
The field is constrained by limited sMRI datasets, unavailable functional MRI datasets, and insufficient hardware resources with associated real-world privacy concerns.
Abstract
from arXiv · showhide
Multiple Sclerosis (MS) is a type of brain disease which causes visual, sensory, and motor problems for people with a detrimental effect on the functioning of the nervous system. In order to diagnose MS, multiple screening methods have been proposed so far; among them, magnetic resonance imaging (MRI) has received considerable attention among physicians. MRI modalities provide physicians with fundamental information about the structure and function of the brain, which is crucial for the rapid diagnosis of MS lesions. Diagnosing MS using MRI is time-consuming, tedious, and prone to manual errors. Hence, computer aided diagnosis systems (CADS) based on artificial intelligence (AI) methods have been proposed in recent years for accurate diagnosis of MS using MRI neuroimaging modalities. In the AI field, automated MS diagnosis is being conducted using (i) conventional machine learning and (ii) deep learning (DL) techniques. The conventional machine learning approach is based on feature extraction and selection by trial and error. In DL, these steps are performed by the DL model itself. In this paper, a complete review of automated MS diagnosis methods performed using DL techniques with MRI neuroimaging modalities are discussed. Also, each work is thoroughly reviewed and discussed. Finally, the most important challenges and future directions in the automated MS diagnosis using DL techniques coupled with MRI modalities are presented in detail.
1. Introduction
MS damages the central nervous system and can produce diverse symptoms, while MRI-based diagnosis is clinically valuable but time-consuming and challenging. This review surveys deep-learning CAD systems for automated MS detection using MRI modalities, alongside datasets, methods, challenges, and future directions.
- MS is a chronic autoimmune disease in which the immune system damages the central nervous system, including the brain and spinal cord.
- 2.3 million people worldwide were estimated to have MS from 2013 onward, and the disease can cause symptoms including fatigue, mobility difficulty, weakness, and vision problems.
- MRI is described as the best non-invasive diagnostic method, revealing lesions where MS-related myelin damage increases water content.
- MRI-based MS diagnosis is time-consuming and challenging for physicians, motivating AI methods that support automated identification of disease-related patterns.
- Deep learning integrates feature extraction and classification within CAD systems, avoiding the manual feature-management required by conventional machine learning.
- The review covers DL-based MS lesion segmentation and detection, using MRI datasets and architectures, and discusses research trends, challenges, and future directions.
(5). Intensity standardization
MRI intensity patterns vary across scanners and scanning sessions, so preprocessing standardizes images before subsequent analysis.
- MRI intensity patterns can differ across scanners and sessions even under the same acquisition protocol.
- Histogram matching is the most popular procedure described for correcting scanner-dependent intensity variations.
- De-oblique preprocessing is used before registration because oblique scanning complicates alignment between MR images.
- Re-orientation techniques enforce identical image directions to reduce misregistration caused by directional differences.
(8). Segmentation
Segmentation isolates meaningful tissue or lesion regions in MRI and is central to DL-based MS diagnosis, where preprocessing and CNN architectures support lesion delineation and classification.
- MRI segmentation isolates white matter, gray matter, and cerebrospinal fluid according to shared intensity, depth, color, or structural characteristics.
- High- and low-level preprocessing methods, including data augmentation, patch extraction, and region-of-interest extraction, are used to improve CAD performance.
- Most reviewed DL studies focus on segmentation, classification, or both, with CNNs, autoencoders, GANs, and CNN-RNN models represented.
- CNN models avoid manual feature extraction and learn higher-level features as network depth increases.
- The review discusses 2D and 3D CNNs for MS classification and U-Net and FCN architectures for MRI lesion segmentation.
(1). 2D and 3D-CNNs
CNNs are widely used for MRI segmentation and classification because they reuse weights and reduce parameter counts for 2D and 3D images, although deeper models increase computational demands.
- CNNs support image segmentation and classification and are compatible with 2D and 3D images through parameter reduction and weight reuse.
- Convolutional, pooling, batch-normalization, and fully connected layers are identified as key CNN components.
- A 2D-CNN block diagram is presented for automated MS detection.
(2). Pre-Trained CNN Networks
Pre-trained CNNs address the difficulty of training deep learning networks from scratch on small MRI datasets by adapting models initially trained on large datasets.
- Small datasets may produce poor classification when deep learning networks with many training parameters are trained from scratch.
- Transfer learning replaces a pre-trained model’s classification layer with a problem-specific layer before updating weights using new data.
- Popular pre-trained models for MRI-based MS diagnosis include LeNet, AlexNet, GoogleNet, VGGNet, and ResNet.
- Using ImageNet for initial training is disadvantageous because MRI images are grayscale.
(3). FCN Network
The reviewed section describes deep learning architectures used for automated MS diagnosis with MRI, including FCN, U-Net, GAN, autoencoder, and CNN-RNN designs. These models support segmentation, classification, representation learning, and related MRI analysis tasks.
- FCN Network: Fully convolutional networks replace fully connected layers with convolution layers and up-sample local output maps for image segmentation.
- U-Net: U-Net combines encoder down-sampling and decoder up-sampling paths with skip connections between corresponding layers.The architecture is described as suitable for image segmentation and trainable with limited data.
- GAN: Generative adversarial networks use a generator to produce samples and a discriminator to distinguish original from generated data.The two networks compete during training; gradient vanishing and training complexity are identified as limitations.
- Autoencoder: Autoencoders compress high-dimensional inputs into a latent-space representation and decode them back to the original dimensions.
- CNN-RNN: CNN-RNN architectures combine CNN-based spatial-feature learning with RNN-based temporal-feature learning.The CNN output feature maps are passed to an RNN after convolutional processing.
3. Discussion
The review compares deep-learning applications, datasets, MRI modalities, preprocessing tools, network architectures, toolboxes, and classification components used for automated MS diagnosis. Segmentation and classification dominate applications, while sMRI, CNNs, FSL, and Keras are reported as especially common choices.
- Applications: Segmentation and classification, alone or combined, dominate deep-learning applications for MS diagnosis and help identify lesion location and dimensions.Manual lesion-segmentation datasets have supported the growth of segmentation studies.
- Datasets: Clinical datasets are used most often, with ISBI 2015 reported as the most frequently used dataset for MS diagnosis research.ISBI 2015 contains multiple sMRI modalities, motivating its use in many studies.
- MRI modalities: sMRI use has grown more than other neuroimaging modalities, while a few studies combine PET with sMRI.The review associates this combination with improved CADS precision and efficiency.
- Preprocessing: FSL is the most widely used MRI preprocessing toolbox among the reviewed studies.Other named preprocessing tools include FreeSurfer, SPM, and Matlab.
- DL architectures: CNNs are the most widely used deep-learning networks for MS diagnosis from MRI, partly because they achieve high performance on brain MR images.CNNs also extract higher-level features without manual feature extraction, although deeper models increase computational complexity.
- DL tools and classification: Keras is the most used deep-learning toolbox, while softmax has yielded the highest classification performance among reported final-layer activation functions.The review describes Keras as free, open-source, and easy to use.
4. Challenges
The review identifies limited data access and constrained computational resources as major challenges for automated MS diagnosis using MRI and deep learning. These constraints affect sMRI coverage, functional neuroimaging research, and the feasibility of advanced models.
- Challenges: The main challenges are inaccessible large sMRI databases with diverse subjects and modalities, inaccessible functional-neuroimaging datasets, and limited DL models or hardware resources.The review presents these as three central challenges for accurate automated MS diagnosis.
- sMRI data: Available datasets contain finite subject numbers, limiting use of advanced deep-learning models that require huge datasets for high classification performance.Segmentation additionally requires many manually delineated MR images from large numbers of subjects.
- Functional neuroimaging: Functional-neuroimaging research is constrained by the inaccessibility of fMRI and other functional-neuroimaging datasets.The review identifies dataset access as the principal challenge for these modalities.
(1). Unavailable fMRI datasets
The review identifies the unavailability of functional MRI datasets as a key challenge for automated MS diagnosis. Although fMRI can provide information about brain function and lesion location, its greater complexity has limited access and use.
- (1). Unavailable fMRI datasets: fMRI datasets are unavailable, limiting research on automated MS diagnosis.The review identifies this as a key challenge.
- (1). Unavailable fMRI datasets: fMRI includes resting-state and task-based categories and provides information about brain function.These modalities can help specialists examine motor, visual, and cognitive networks in people with MS.
- (1). Unavailable fMRI datasets: fMRI can determine MS lesion locations based on the functioning of brain neurons, but it is more complex than structural MRI.This complexity has contributed to limited research using fMRI and AI for MS detection.
(2). Unavailable other functional neuroimaging datasets
Access to datasets from other functional neuroimaging modalities is also limited, while multimodality studies for MS remain scarce. The review notes that data fusion could improve system performance.
- (2). Unavailable other functional neuroimaging datasets: Datasets from EEG, fNIRS, and MEG are difficult to access, and only a few clinical studies have used them for MS diagnosis.The review also notes limited use of EEG with conventional machine learning.
- (2). Unavailable other functional neuroimaging datasets: Very few clinical studies have diagnosed MS using multimodality combinations such as EEG-fMRI and MEG-fMRI.The review identifies multimodality diagnosis as an underexplored area.
- (2). Unavailable other functional neuroimaging datasets: Data fusion could improve the performance of systems that combine multiple neuroimaging modalities.The passage presents this as a potential benefit of multimodality approaches.
5. Future directions in the automated MS diagnosis using DL techniques
Future work emphasizes expanding datasets, developing newer deep learning models, and applying cloud-based technologies to automated MS diagnosis. Proposed directions address limited data, model development, and computational access.
- Future directions: Future directions are grouped into datasets, novel deep learning models, and rehabilitation systems for people with MS.These categories define the review's overall organization of proposed research directions.
- Datasets: Future research should develop larger structural MRI datasets and more accessible fMRI, PET, and combined PET-MRI datasets.The review links small structural MRI datasets and limited functional-modality access to challenges in building practical systems.
- Novel deep learning models: Generative adversarial networks could generate additional MRI data to address limited medical datasets for deep learning.The review also proposes graph-based architectures, zero-shot learning, and representation learning as future approaches.
- Technology-enabled research: Cloud computing could support storage, implementation, and simulation of deep learning methods for MS diagnosis using MRI.The review expects future studies to use cloud-based resources in this area.
6. Conclusion
The review surveys deep learning-based CAD systems for automated MS diagnosis with MRI, covering datasets, modalities, preprocessing, architectures, and classifiers. It also emphasizes challenges involving dataset scale and diversity, access to functional imaging, and practical deployment resources.
- 6. Conclusion: Deep learning-based CAD systems are reviewed for automated MS lesion segmentation and detection using MRI modalities.The review discusses the components and automated systems used for MS diagnosis.
- 6. Conclusion: The review compares conventional machine learning and deep learning alongside MRI datasets, modalities, preprocessing tools, models, toolboxes, and classifiers.These elements form the main comparison scope described in the conclusion.
- 6. Conclusion: Limited access to large structural MRI datasets from diverse populations remains an important challenge for automated MS diagnosis.The conclusion identifies dataset scale and population diversity as unresolved constraints.