AI-assisted longitudinal MRI assessment in multiple sclerosis: diagnostic performance and reduction in reporting time.
Authors
Affiliations (6)
Affiliations (6)
- Neuroradiology Department, IRCCS Ospedale San Raffaele, Milan, Italy. [email protected].
- Neuroradiology Department, IRCCS Ospedale San Raffaele, Milan, Italy.
- Neuroimaging Research Unit, Institute of Experimental Neuroscience, IRCCS Ospedale San Raffaele, Milan, Italy.
- Vita-Salute San Raffaele University, Milan, Italy.
- Neuroradiology Department, IRCCS Ospedale San Raffaele, Milan, Italy. [email protected].
- Vita-Salute San Raffaele University, Milan, Italy. [email protected].
Abstract
Detection of new and enlarging lesions on longitudinal MRI is essential for monitoring disease activity in multiple sclerosis (MS) but is time-consuming and subject to inter-reader variability. This study evaluated the diagnostic performance and workflow impact of a CE-marked artificial intelligence (AI) platform compared with neuroradiologists of different experience. In this retrospective study, 59 patients with relapsing-remitting MS underwent baseline and follow-up brain MRI including 3D FLAIR sequences. Two neuroradiologists independently assessed examinations for new or enlarging lesions, blinded to AI output. The reference standard was an AI-blinded consensus of the two readers, with discordant cases adjudicated by an independent third neuroradiologist. Patient-level diagnostic performance was evaluated for each reader alone, the AI alone, and each reader assisted by AI. Agreement was assessed with Cohen's kappa and reading times with the Wilcoxon signed-rank test. Among unassisted assessments, the experienced reader achieved the highest accuracy (88.1%), followed by the less experienced reader (83.1%) and the AI alone (79.7%). AI assistance was associated with higher accuracy for both readers (94.9%, κ = 0.90, and 89.8%, κ = 0.80, respectively). AI assistance significantly reduced reading time for both readers (median within-patient reduction 0.9 and 1.0 min; p < 0.001), independently of baseline T2 lesion volume. Against an AI-blinded consensus reference standard, AI assistance was associated with higher diagnostic performance of both readers, irrespective of experience, and significantly reduced reading time, supporting its role as a decision-support tool in the longitudinal MRI assessment of MS.