'Can Deep Learning-Based MRI Techniques Improve the Accuracy and Efficiency of Multiple Sclerosis (MS) Detection Compared to Conventional MRI Methods? Systematic Review'

Abstract

Background: Multiple sclerosis (MS) is a chronic neuroinflammatory disorder and a leading cause of non-traumatic neurological disability among young adults. Early and accurate diagnosis is critical for initiating disease-modifying treatments and improving long-term outcomes. Conventionally, MRI enhanced with gadolinium-based contrast agents (GBCAs) plays a key role in identifying active lesions; however, concerns about safety, cost, and accessibility limit its widespread use. With recent advancements in artificial intelligence (Al), especially deep learning (DL), there has been growing interest in contrast-free diagnostic solutions based on non-contrast MRI. Nevertheless, questions remain regarding the diagnostic accuracy and clinical viability of these emerging approaches. Aims: This systematic review aimed to evaluate whether DL-based MRI methods using non-contrast sequences can effectively detect MS lesions and match or outperform conventional contrast-enhanced MRI in terms of diagnostic accuracy and efficiency. Methods: A comprehensive search of medical databases was conducted to identify peer-reviewed studies evaluating DL models for MS diagnosis using non-contrast MRI. Following rigorous screening and quality assessment using the QUADAS-2 tool, nine studies were included-five utilizing conventional MRI with GBCAs and four applying DL-based methods on non-contrast imaging. Metrics such as Dice Similarity Coefficient (DSC), sensitivity, specificity, and inter-rater reliability were extracted and analyzed narratively. Results: DL models trained on non-contrast sequences achieved strong lesion segmentation accuracy, with DSCs ranging from 0.74 to 0.96 and high inter-rater agreement (ICC > 0.87). While conventional MRI studies reported high sensitivity (up to 100%) for detecting active lesions, DL-based approaches demonstrated comparable performance, particularly in cortical and juxtacortical regions-without requiring contrast agents. Conclusion: The included studies demonstrate that DL-based MRI has strong potential to serve as a reliable, contrast-free diagnostic alternative in MS imaging. Despite the promising performance metrics, wider adoption will depend on external validation, larger sample sizes, and standardized evaluation protocols. As evidence accumulates, DL models may soon play a pivotal role in reshaping the landscape of MS diagnosis through safer and more accessible imaging tools.