AI-Assisted Follow-Up MRI Interpretation in Multiple Sclerosis: Radiology Residents’ Diagnostic Confidence and Performance
Year: 2026
Type: article
Abstract: Background/Objectives: Artificial intelligence (AI) tools are becoming increasingly utilised in everyday radiological practice. Evaluating follow-up magnetic resonance imaging (MRI) of patients with multiple sclerosis (MS) is challenging. The aim of this study was to assess changes in confidence among four participating radiology residents when evaluating follow-up MRIs of MS patients using Pixyl.Neuro.MS® software. Methods: This retrospective study included 78 MS patients. Previous and new FLAIR images were independently analysed by four residents, first without AI and then in an immediate AI-assisted second reading using Pixyl.Neuro.MS®, a deep-learning lesion-segmentation software. Readers recorded their findings and confidence after each reading and rated the impact and usefulness of AI. The reference standard was defined by a consensus reading of a neuroradiologist and a radiology resident (15 and 4 years of experience). Paired analyses and patient-level bootstrap confidence intervals were used; bootstrap inference was conditional on the four readers. Results: The reference standard identified 24 patients with new lesions. Mean accuracy across readers increased from 84.62% to 90.71% (difference, 6.09 percentage points; 95% bootstrap confidence interval [CI], 2.24–10.26). Only Reader 3 showed a significant accuracy increase after Holm adjustment (p = 0.025). No accuracy comparison with routine reports was significant after adjustment. Confidence increased for all readers (all Holm-adjusted p < 0.001). Perceived usefulness did not differ significantly between lesion-burden groups. All readers predominantly rated the impact of AI as positive. The second-reading phase was shorter by a mean of 59.35 s (95% bootstrap CI, 51.39–67.07). Conclusions: In these four residents, the AI-assisted second reading was associated with higher confidence and average accuracy. The immediate sequential design prevents separation of AI contributions from familiarity and order effects, and the shorter second-reading phase does not establish an overall workflow time saving.
Source: Diagnostics
Institutions: Centre National de la Recherche Scientifique, Inserm, University of Novi Sad, Commissariat à l'Énergie Atomique et aux Énergies Alternatives, Université Paris-Saclay, CEA Paris-Saclay, Laboratoire d'Imagerie Biomédicale Multimodale Paris-Saclay, Arts et Métiers, Université Sorbonne Paris Nord, Institut de Biomécanique Humaine Georges Charpak, Assistance Publique – Hôpitaux de Paris, Oncology Institute of Vojvodina, Observatoire Français de la Sclérose en Plaques, Centre Hospitalier Intercommunal de Poissy
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: gold
Estimated APC paid: $3,137
OpenAlex ID: W7235892162
