Loredana Storelli1, Elisabetta Pagani1, Maria Assunta Rocca1,2, Paolo Preziosa1,2, Antonio Gallo3,4, Gioacchino Tedeschi3,4, Maria Laura Stromillo5, Nicola De Stefano5, Hugo Vrenken6, David Thomas7, Laura Mancini7, Christian Enzinger8, Franz Fazekas8, and Massimo Filippi1,2

1Neuroimaging Research Unit, San Raffaele Scientific Institute, Vita-Salute San Raffaele University, Milan, Italy, 2Department of Neurology, San Raffaele Scientific Institute, Vita-Salute San Raffaele University, Milan, Italy, 3MRI Center “SUN-FISM”, Second University of Naples and Institute of Diagnosis and Care “Hermitage-Capodimonte, Naples, Italy, 4I Division of Neurology, Department of Medical, Surgical, Neurological, Metabolic and Aging Sciences, Second University of Naples, Naples, Italy, 5Department of Neurological and Behavioral Sciences, University of Siena, Siena, Italy, 6Department of Radiology and Nuclear Medicine, MS Centre Amsterdam, VU University Medical Centre, Amsterdam, Netherlands, 7NMR Research Unit, Queen Square MS Centre, UCL Institute of Neurology, London, United Kingdom, 8Department of Neurology, Medical University of Graz, Graz, Austria


Aim of the study was to develop a semi-automatic method for the segmentation of hyperintense multiple sclerosis (MS) lesions on dual-echo (DE) PD/T2-weighted scans. DE MRI scans were obtained from 6 different European centers from 52 MS patients with a mean lesion load of 10.3 (± 11.9) ml. The method was based on a region growing approach initialized by manual identification of lesions and a priori information. The segmentation results with the new method showed high accordance with the ground truth and a low misclassification of lesion voxels. Furthermore, operator time required for lesion segmentation was drastically reduced.


To develop a semi-automatic method for the segmentation of hyperintense lesions on dual-echo (DE) PD/T2-weighted magnetic resonance images (MRI) from patients with multiple sclerosis (MS).


The analysis of disease burden on MRI from MS patients, both for research and clinical trials, requires the quantification of the volume of hyperintense lesions on a T2-weighted MRI sequence (1). Several automatic methods for MS lesion segmentation have been proposed, which have not been trained or validated on DE datasets and may thus identify lesions false positively and false negatively (2). Manual segmentation still remains the gold standard although it is time-consuming and introduces inter- and intra-observer variability.


DE MRI scans were obtained using 3 different 3 Tesla scanners (GE, PHILIPS, SIEMENS) from 52 MS patients from 6 different centers (Amsterdam, Graz, London, Milan, Naples and Siena) with a mean lesion load of 10.3 (± 11.9) ml. PD-w image intensity values were standardized to correct for the arbitrary intensity scaling for different acquisitions and scanners (3), thus the intensity histogram of each given image was rescaled to the standard one. The core of the algorithm was the pixel-based region growing segmentation method. This approach examines neighbouring pixels of initial "seed points" and determines whether the pixel neighbours should be added to the region according to similarity constraints (4). Thus, starting from the seed point (manually identified by a single expert physician using the DE dataset, cf fig 4), the expansion of the segmented region continued to the adjacent pixels until for all neighbouring pixels the stop condition was reached. This condition was a combination of two constraints: 1) the intensity of the new voxel was dissimilar to that of the seed point, according to a function determined after a training process on manual segmentation (Figure 1); 2) a lesion edge was reached, based on the high-pass filtering of a half-way contrast image that was obtained by averaging the not standardized PD-w and T2-w images to take advantage of both images´ tissue contrasts (Figure 2). After an initial segmentation, a more robust intensity threshold was estimated, considering the distribution of intensity values for each segmented lesion. This new threshold was used to restart the region growing for each lesion and refine the initial segmentation. The algorithm (Figure 3) was implemented in Matlab®. Manual segmentations by an expert operator on XY sequences were used as the gold standard (Figure 4). The metrics evaluated were Dice Similarity Coefficient (DSC), Root Mean Square Error of lesion load (RMSE), Sensitivity, True Positive Fraction (TPF), False Positive Fraction (FPF), and False Negative Fraction (FNF) for each patient (Figure 5).


The following validation measures averaged over all patients were obtained: DSC = 62%; RMSE = 2 ml; TPF = 0.76; FPF = 0.36; FNF = 0.22. For the considered lesion loads, the average time for manual lesion segmentation was about 50 minutes, while for the new method the average time was about 55 seconds for one patient.


Lesion segmentation performed using the new method was very similar to the ground truth. FPF and FNF values indicated low misclassification of lesion voxels. Moreover, the operator time required to extract lesion volumes was drastically reduced.


This study was partially supported by Fondazione Italiana Sclerosi Multiple (FISM2013/S/1). Data were collected within the MAGNIMS network.


(1) Filippi M, Rocca M.A, De Stefano N, Enzinger C, Fisher E, Horsfield M.A, Inglese M, Pelletier D, Comi G, Magnetic resonance techniques in multiple sclerosis: the present and the future. Arch Neurol. 68(12):1514-20. doi: 10.1001/archneurol.2011.914. Review; 2011.

(2) Garcia-Lorenzo D., Francis S., Narayanan S., Arnold D. L., Collins D. L. Review of automatic segmentation methods of multiple sclerosis white matter lesions on conventional magnetic resonance imaging. Med Image Anal, 17, 1-18; 2013.

(3) Nyul, L. G. and Udupa, J. K., On standardizing the MR image intensity scale. Magn Reson Med, 42, 1072-1081; 1999.

(4) Kamdi S. and Krishna R.K.. Image Segmentation and Region Growing Algorithm. International Journal of Computer Technology and Electronics Engineering, 2, 103-107; 2012.


Figure 1. Threshold values extracted after the training process on the manually segmented lesions. The red line is the fitted line used to select the threshold function for the region growing approach.

Figure 2. An example half-way contrast image (c) obtained by averaging the not standardized PD-w image (a) and the T2-w image (b).

Figure 3. Schematic overview of the implemented method.

Figure 4. Examples of lesion segmentation for three different patients (in the three rows) performed by an expert operator (in blue) compared to the performance of the proposed method (in red). The corresponding T2-w images are shown in the right column.

Figure 5. In the top left graph, DSC values are reported and in the top right graph, the mean TPF, FPF and FNF values are shown for each patient. In the bottom plots, sensitivity and a scatter plot to compare manual lesion load against automatic lesion load are shown.

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)