AUTHORS: La Rosa F, Joao Fartaria M, Abdulkadir A, Rahmanzadeh R, Lu PJ, Galbusera R, Granziera C, Thiran JP, Bach Cuadra M

Multiple Sclerosis Journal, 25(7): 131-356, June 2019

ABSTRACT

Background and aim:

 Since 2017, Multiple Sclerosis (MS) cortical lesions (CL) are included in the McDonald diagnostic criteria of MS and evidence suggests that they more accurately determine patients’ cognitive deficits than white matter (WM) lesions. Advanced Magnetic Resonance Imaging (MRI) sequences, such as Double Inversion Recovery (DIR), and Magnetisation Prepared 2 Rapid Acquisition Gradient Echoes (MP2RAGE), provide enough contrast to visualise CL; however, their detection and segmentation is reader-dependent and tedious.
We aimed at developing a deep learning approach based on a convolutional neural network (CNN) to automatically detect CL on DIR, Fluid-Attenuated Inversion Recovery (FLAIR), and MP2RAGE images. Furthermore, we evaluate the performance of the CNN using FLAIR and MP2RAGE sequences only.

Methods: 

We propose a novel patch-based CNN consisting of convolutional and max-pooling layers. Input patches have size 15x15x15 voxels and are centered on a lesion candidate, whereas the output patch size is 3x3x3 voxels. The network is trained with a weighted loss function consisting of a combination of the Tversky index and cross-entropy. In this way, both WM lesions and CL are exploited during the training phase, but the CNN learns to focus more on CL. The network is evaluated on two datasets of MS patients. The first one consists of 36 cases with FLAIR, DIR and MP2RAGE contrasts. The second one includes 35 cases with only FLAIR and MP2RAGE sequences. In both datasets MS lesions were manually detected and delineated by an expert radiologist.

Results: 

We evaluated the CNN on two different scenarios. First, we divided dataset 1, which includes all three MRI contrasts, in training, validation, and testing (20, 2, 14 subjects respectively). Second, we considered the FLAIR and MP2RAGE sequences only. In this case the network was trained on dataset 1 (36 subjects) and tested on dataset 2 (35 subjects). We obtained a 75% CL, 63% WM detection rate and 32% lesion-wise false positives rate (LFPR) on the first scenario, and 58% CL, 39% WM lesion detection rate, and 38% LFPR on the second one.

Conclusion: 

To the best of our knowledge, we present the first deep learning-based approach for an automatic detection of CL in MS patients using combinations of MRI contrasts, achieving a 75% detection rate in the best scenario. DIR seems to be essential in CL detection, without this, our CNN achieved a 58% CL detection rate, despite the larger training dataset.

BibTex

Module: