AUTHORS: Todea RA, Melie Garcia L, Barakovic M, Siebenborn N, Cagol A, Rahmanzadeh R, Galbusera R, Lu PJ, Weigel M, Ruberte E, Radue EW, Schaedelin S, Benkert P, Yaldizli O, Oechtering J, Sinnecker T, Müller S, Achtnichts L, Vehoff J, Disanto G, Findling O, Chan A, Salmen A, Pot C, Lalive P, Zecca C, Derfuss T, Remonda L, Wagner F, Vargas M, Du Pasquier R, Pravata E, Weber J, Gobbi C, Leppert D, Wuerfel J, Kober T, Marechal B, Corredor-Jerez R, Psychogios M, Lieb J, Kappos L, Bach Cuadra M, Kuhle J, Granziera C

ECTRIMS 2021 - Multiple Sclerosis Journal, 27(2_SUPPL): 499, October 2021


Introduction: The LeMan-PV is an automated tool allowing longitudinal segmentation of MS lesions, based on a combination of lesion-segmentation and change-detection approaches. To date, the prototype has been validated only in one monocentric study.

Objectives: To assess the diagnostic accuracy of LeMan-PV for the detection of new and enlarging MS lesions in a multicentric longitudinal study.

Methods: We retrospectively studied 206 MS patients (313MRI pairs, 626 scans, 123 at 3T, and 190 at 1.5T) from five centers participating in the Swiss Multiple Sclerosis Cohort study (SMSC). The identification of new lesions was performed using LeMan-PV on 3D-FLAIR and MPRAGE images. A lesion was considered enlarged by LeMan-PV when it showed an increase of at least 10 voxels. A neuroradiologist detected manually the new and enlarging lesions according to standard clinical guidelines. Neuroradiologically, a lesion was considered enlarged when it showed an enlargement of 3mm or more in at least 2 planes and new when it was not present at the baseline. We then assessed the sensitivity, specificity, and F1 score of LeMan-PV vs. the manual reading (i) in the whole cohort and (ii) and separately in scans acquired at 1.5T or 3T.

Results: LeMan-PV detected new lesions with a sensitivity (Sens) of 73%, specificity (Spec) of 71%, accuracy (Acc) of 72% and an F1 score (F1s) of 0.46. The performance was lower for enlarging lesions (Sens:72%, Spec: 21%, Acc:25%, F1s: 0.11).

When analysed only images acquired at 3T (2 centers), LeMan-PV was overall more accurate for new lesions (Sens: 87%, Spec:74%, Acc: 76% and F1s:0.34) but not for enlarged lesions (Sens: 72%, Spec: 21%, Acc: 25%, and F1s: 0.11).

When analysed only images acquired at 1.5T (4 centers), LeMan-PV showed overall lower performance than at 3T (new lesions, Sens: 70%, Spec: 70%, Acc: 70%, F1s:0.51; enlarged lesions, Sens: 72%, Spec: 18%, Acc: 23%, F1s: 0.15).

Some centers reached more accurate automated detection than others, with the best center showing a Sens: 88%, Spec: 90%, Acc: 89%, and F1s:0.8).

Discussion and Conclusions: LeMan-PV performed better with new lesions than with enlarging lesions, probably due to the different thresholds to enlargement applied by the software and the standard clinical reading. The multicentric performance of LeMan-PV was better than that of other longitudinal tools that were tested in multi-centric studies.