Magnetic Resonance Imaging (MRI) plays a major role in Multiple Sclerosis (MS) diagnosis, follow-up and therapy monitoring. More specifically, the identification of new or resolved lesions and changes in lesion size due to the progression of inflammation and demyelination is important to perform early diagnosis¹, quantify ongoing disease activity and monitor treatment effects^2,3. In this work, we assess the performance of an in-house automated tool⁴, to detect and segment longitudinal changes in MS lesions based on a clinical and an "advanced" MRI protocol⁵.

3T MR images were acquired on a MAGNETOM Trio a Tim system (Siemens Healthcare, Germany) using a 32-channel head coil. Thirty-two patients with relapsing-remitting MS and disease duration <5 years from diagnosis were enrolled in the study, and two MRIs were performed at enrolment (TP1) and at two-years (21.4 ± 2.5 months, range 16-27 months) follow-up (TP2). The patient cohort consisted of 13 males and 19 females, age range 20-60 years at TP1, with a median Expand Disability Status Scale (EDSS) of 1.5 at both time points. The MRI protocol included:

- Magnetization-Prepared Rapid Acquisition Gradient Echo (MPRAGE, TR/TI=2300/900ms, voxel size (vs)=1.0x1.0x1.2mm³);

- Magnetization-Prepared 2 Rapid Acquisitions Gradient Echo (MP2RAGE, TR/TI1/TI2=5000/700/2500 ms, vs=1.0x1.0x1.2mm³);

- 3D FLuid-Attenuated Inversion Recovery (FLAIR, TR/TE/TI=5000/394/1800, vs=1.0x1.0x1.2mm³);

- 3D Double Inversion Recovery (DIR, TR/TE/TI1/TI2=10000/218/450/3650, vs=1.1x1.0x1.2mm³).

At both time points, a supervised classifier based on the k-nearest-neighbour (k-NN) algorithm was used to determine the lesion probability of each image voxel using the following features: 1)image intensity in each contrast, 2)spatial coordinates in a reference space⁶, and 3)tissue prior probabilities⁷. Manual detection of MS lesions (from a neurologist and a radiologist) was used as ground truth (GT) for both time points as well as to train the classifier. Minimum lesion size was set at 0.009 mL⁸. Grey- and white-matter (GM, WM) lesions were classified in 5 groups with the following criteria⁹:

(i)new: identifiable on the registered TP2 image but not on the registered TP1 image;

(ii)enlarged: increase in diameter by at least 50%;

(iii)resolved: clearly visible on the registered TP1 image but not on the registered TP2 image;

(iv)shrunken: decrease in diameter by at least 50%;

(v)unaltered: do not follow any of the above criteria. The performance was evaluated through a “leave-one-out” cross-validation. Detection rate (DR, number of detected lesions/total GT lesions) per brain for the different types of lesions (i-v) was computed using: (1)clinical images (FLAIR and MPRAGE) and (2)clinical and research images (FLAIR, MP2RAGE and DIR, “advanced protocol”). The confusion table of average detection rate per lesion type was computed. Lesion detection performance was compared between (1) and (2) using the Wilcoxon signed-rank test. Automated estimation of total lesion volume difference between TP1 and TP2 per patient was evaluated through a Bland-Altman plot¹⁰.

Detection rate for all types of lesions improved significantly when the "advanced protocol" (DR=81.3%) was used instead of the “clinical protocol” (DR=78.7%, p<0.01, Figure 2). Statistical differences in DR were short of significance for particular lesion types, except for lesion type (iii) (resolved lesions, p<0.05). However, the best median detection rates in all lesion types were consistently obtained using the "advanced protocol" (Figure 2): (i) 62.5%, (ii) 100%, (iii) 66.7%, (iv) 100%, and (v) 89.4%. Misclassification of different types of lesions was also significantly improved for resolved and unaltered lesions using the "advanced protocol" (p<0.05, Figure 3). Except for three patients, volume difference quantification lies within ±1.96 standard deviations, indicating the good agreement between manual and automated segmentations using any type of protocol(Figure 4).The method exhibits good performance in detection of enlarged, resolved and unaltered lesions. New and resolved lesions present the lowest DR compared to the other lesion types, possibly due to lower size (strongly affected by partial volume) and low contrast (low degree of tissue inflammation/demyelination). However, in some cases, new/resolved lesions are detected as enlarged, shrunken or unaltered in both time points, which was retrospectively confirmed as correct by an expert. This highlights the difficulty in distinguishing focal lesions from diffuse damage, even by experts, and demonstrates the ability of the algorithm to detect lesions in both time points that were missed by manual segmentation. When the "advanced protocol" is used, the variability of lesion volume differences(TP1-TP2) between manual and automated segmentation is lower. This can be explained by the higher sensitivity in detection of cortical lesions and the better lesion delineation when advanced sequences are used, as shown in previous work⁴. Future work will aim to reduce partial-volume effects and to optimize the weighting of contrasts in order to increase the sensitivity of both GM and WM lesion detection.