Evaluation of two AI techniques for the detection of new T2/FLAIR lesions in the follow-up of multiple sclerosis patients

Milica Mastilovic^1,2Olivier Heinzlef^3,4Christian Federau^5,6Verónica Muños-Ramírez⁷Marie Blanchere^3,4Jasmina Boban^2,8Francois Cotton^10,9Myriam Edjlali^1,11*

• ¹Laboratoire d'imagerie Biomédicale Multimodale (BioMaps), Université Paris-Saclay, CEA, CNRS, Inserm, Service Hopsitalier Frédéric Joliot, Orsay, France
• ²Faculty of Medicine, University of Novi Sad, Novi Sad, Serbia
• ³Department of Neurology, Poissy-Saint-Germain-en-Laye Hospital, Poissy, France
• ⁴CRC SEP IDF Ouest, Poissy-Garches, France
• ⁵AI Medical AG, Zollikon, Switzerland
• ⁶Universitat Zurich, Zürich, Switzerland
• ⁷Pixyl Research and Development Laboratory, Grenoble, France
• ⁸Oncology Institute of Vojvodina, Sremska Kamenica, Serbia
• ⁹Radiology Department, Centre Hospitalier Lyon-Sud, Hospices Civils de Lyon, Oullins-Pierre-Bénite, France
• ¹⁰CREATIS, INSERM U1044, CNRS UMR 5220, UCBL1, Villeurbanne, France
• ¹¹Department of Radiology, APHP, Hôpitaux Raymond-Poincaré & Ambroise Paré, Paris, France

Background: Multiple sclerosis is an inflammatory demyelinating disease of the CNS. Annual MRI exams are crucial for disease monitoring. Interpreting high T2/FLAIR lesion loads can be laborious. AI aids in lesion detection, and choosing between different solutions can be challenging. Aim: This study compares two distinct software, Pixyl.Neuro.MS® and Jazz®, to assess their performance in T2/FLAIR lesion detection between two-time points. Methods: Retrospective analysis included follow-up MRIs from 35 MS patients. Pixyl.Neuro.MS® automatically segments and classifies lesions. Jazz® automates the reading process and image display. Two readers (15 and 4 years of experience) conducted radiological analysis, followed by AI-assisted readings. A number of new lesions (NL) and reading times were recorded, with ground truth (GT) established by consensus. AI-detected lesions were classified as true (TP) and false positives (FP). Statistical analysis used SPSS (p<0.05). Results: Pixyl.Neuro.MS® readings averaged 2 minutes 46 seconds ± 1 minute 4 seconds while using Jazz® 3 minutes 33 seconds ± 2 minutes 24 seconds. Over 50% of the population had a high lesion load (>20 lesions). Both software significantly improved NL detection (p<0.01 for both), revealing them in more patients than standard readings. Standard reports found 8 NL in 2 patients, while AI-assisted readings detected at least 17 TP in 7 patients and rejected 61 FP lesions. GT detected 21 lesions in 19 patients. Conclusions: Both AI software have been found to enhance NL detection in MS patients, outperforming standard methods. These tools offer crucial advantages for accurate disease monitoring.