مقالات پذیرفته شده در نهمین کنگره بین المللی زیست پزشکی
Diagnostic Accuracy of Convolutional Neural Networks Models for Early Detection of Multiple Sclerosis in MRI Images: A Systematic Review of Imaging-Based Clinical Applications
Diagnostic Accuracy of Convolutional Neural Networks Models for Early Detection of Multiple Sclerosis in MRI Images: A Systematic Review of Imaging-Based Clinical Applications
Kimia Heydariyar,1,*Abdollah Karimi,2Alireza Pourrahim,3Alireza Vasiee,4Omid Raiesi,5Mohammad Mahdi Pourrahim,6
1. Student Research Committee, Faculty of Medicine, Ilam University of Medical Sciences, Ilam, Iran 2. Department of Computer Engineering and Information Technology, Amirkabir University of Technology, Tehran, Iran 3. Student Research Committee, Faculty of Medicine, Ilam University of Medical Sciences, Ilam, Iran 4. Department of Nursing, Faculty of Nursing and Midwifery, Ilam University of Medical Sciences, Ilam, Iran 5. Department of Parasitology, School of Allied Medical Sciences, Ilam University of Medical Sciences, Ilam, Iran 6. Student Research Committee, Faculty of Medicine, Ilam University of Medical Sciences, Ilam, Iran
Introduction: Early diagnosis of Multiple Sclerosis (MS) is critical for effective intervention, yet remains challenging due to subtle radiological presentations. Convolutional Neural Networks (CNNs) offer transformative potential in automating MS detection using MRI. This systematic review synthesizes evidence on the diagnostic accuracy of CNN-based models for early MS identification (clinically isolated syndrome [CIS] or radiologically isolated syndrome [RIS]) via MRI, evaluating their clinical applicability and limitations.
Methods: This PRISMA 2020-compliant review searched PubMed/MEDLINE, IEEE Xplore, Scopus, Web of Science, and arXiv (up to March 2024). MeSH terms included: "Multiple Sclerosis", "Deep Learning", " Convolutional Neural Networks", "Magnetic Resonance Imaging", and "Early Diagnosis".
Keywords combined: (convolutional neural network* OR CNN) AND (multiple sclerosis OR MS OR CIS) AND (MRI OR magnetic resonance imaging) AND (diagnos* OR accuracy OR sensitivity OR specificity).
Eligible studies: 1) Evaluated CNN models for early MS/CIS/RIS diagnosis using MRI; 2) Reported quantitative accuracy metrics (sensitivity/specificity/AUC); 3) Original research. Exclusions: Non-MRI studies, non-CNN models, reviews, or animal studies. Two reviewers independently screened records, extracted data (architecture, dataset size, MRI sequences, performance), and assessed risk of bias using QUADAS-2.
Results: Sixteen studies met inclusion criteria, collectively demonstrating that convolutional neural networks (CNNs) achieved pooled diagnostic accuracy of 90% sensitivity (95% CI: 86–93%) and 92% specificity (95% CI: 88–95%) for early MS detection in MRI, with an aggregate AUC of 0.94 (95% CI: 0.91–0.96). The highest-performing architectures were 3D U-Net models for lesion segmentation (reaching 94% lesion-wise Dice score in CIS cohorts) and ResNet-50/101 variants for classification tasks. Multimodal MRI sequences (T1w+T2w+FLAIR) yielded superior performance (AUC 0.96) compared to single sequences, while spinal cord-specific CNNs achieved 93% sensitivity for early progressive MS. Key limitations included small sample sizes (median *n* = 185; range: 62–1,240), predominant retrospective designs (15/16 studies, 94%), and heterogeneous ground-truth standards (8 distinct lesion annotation protocols). Quality assessment using QUADAS-2 revealed moderate-to-high risk of bias in patient selection (9/16 studies, 56%) due to non-consecutive recruitment and case-control designs, though technical verification bias was low across all studies.
Conclusion: CNNs demonstrate excellent diagnostic accuracy for early MS detection in MRI, outperforming conventional methods. However, clinical translation is impeded by methodological inconsistencies, dataset fragmentation, and lack of prospective validation. Future work must: 1) Standardize MRI acquisition/annotation protocols; 2) Develop explainable CNN architectures; 3) Validate models in multicenter prospective cohorts including diverse populations. While CNNs show promise as diagnostic aids, their integration into clinical workflows requires addressing reproducibility and regulatory hurdles.