Contributions
Abstract: P971
Type: Poster Sessions
Abstract Category: Clinical aspects of MS - Diagnosis and differential diagnosis
Introduction: The central vein sign (CVS), the presence of a vein at the center of white matter (WM) lesions, is a promising imaging marker to differentiate MS from its mimics. Lesions and veins are well contrasted on FLAIR* MRI images, but manually classifying perivenular and non-perivenular lesions is time-consuming.
Objectives: We propose two approaches to automatically detect the CVS: deep learning or a state-of-the-art vessel-filtering approach. We compare their performance and potential for clinical applicability.
Methods: Patients with an established MS diagnosis (n=16) or other MS-mimics (n=17) underwent 3T MRI at the Erasme University Hospital (Brussels, Belgium). Brain WM lesions were manually segmented, and perivenular assessment was done on FLAIR* images by an expert neurologist, following NAIMS guidelines [Sati et al., Nat. Rev. Neurol., 2016; Maggi et al., Ann. Neurol., 2018], yielding 202 perivenular and 212 non-perivenular lesions.
A convolutional neural network (CNN) was designed and trained on 315 3D lesion patches (+36 for validation in each of 10 cross-validation folds) to classify lesions with and without CVS, keeping 63 lesions (5 patients) unseen for final testing. A vesselness filter was also optimized for CVS classification, with average filter response within the lesion mask as the imaging marker. After both algorithms, patients were classified as MS if 50% or more of their lesions were classified as CVS (“50% rule”), as MS-mimic otherwise.
Lesion-level performance was evaluated by specificity, sensitivity, and area under the curve (AUC) on the test set. Patient-level performance was evaluated on the validation set (28 patients) due to the small size of the test set.
Results: Compared to expert visual assessment, at lesion level the CNN reached 81% sensitivity, 71% specificity, and 79% AUC, with near-instant classification (vesselness filter 81%, 57%, 74% resp., around .5 seconds per lesion). At the patient level, the CNN had sensitivity 91%, specificity 88% (vesselness filter 50%, 82% resp.).
Conclusions: The relatively high performance and speed of the proposed automated approaches are promising and show potential for clinical applicability. The CNN performed better than the vesselness filter, without need for lesion masking. Relatively simple improvements to the CNN training and architecture should further increase its performance, which might ease translation of the CVS imaging biomarker into clinical practice.
Disclosure: Jonas Richiardi: Salary paid in part by Siemens Healthcare AG Switzerland
Pietro Maggi: Supported by the ECTRIMS Clinical Training Fellowship Program
Mário João Fartaria: Salary paid by Siemens Healthcare AG Switzerland
Francesco La Rosa: Nothing to disclose
João Jorge: Nothing to disclose
Pascal Sati: Nothing to disclose
Daniel S. Reich: No disclosures relevant to the content of this abstract
Renaud Du Pasquier: Nothing to disclose
Reto Meuli: Nothing to disclose
Meritxell Bach Cuadra: Nothing to disclose
Tobias Kober: Salary paid by Siemens Healthcare AG Switzerland
Abstract: P971
Type: Poster Sessions
Abstract Category: Clinical aspects of MS - Diagnosis and differential diagnosis
Introduction: The central vein sign (CVS), the presence of a vein at the center of white matter (WM) lesions, is a promising imaging marker to differentiate MS from its mimics. Lesions and veins are well contrasted on FLAIR* MRI images, but manually classifying perivenular and non-perivenular lesions is time-consuming.
Objectives: We propose two approaches to automatically detect the CVS: deep learning or a state-of-the-art vessel-filtering approach. We compare their performance and potential for clinical applicability.
Methods: Patients with an established MS diagnosis (n=16) or other MS-mimics (n=17) underwent 3T MRI at the Erasme University Hospital (Brussels, Belgium). Brain WM lesions were manually segmented, and perivenular assessment was done on FLAIR* images by an expert neurologist, following NAIMS guidelines [Sati et al., Nat. Rev. Neurol., 2016; Maggi et al., Ann. Neurol., 2018], yielding 202 perivenular and 212 non-perivenular lesions.
A convolutional neural network (CNN) was designed and trained on 315 3D lesion patches (+36 for validation in each of 10 cross-validation folds) to classify lesions with and without CVS, keeping 63 lesions (5 patients) unseen for final testing. A vesselness filter was also optimized for CVS classification, with average filter response within the lesion mask as the imaging marker. After both algorithms, patients were classified as MS if 50% or more of their lesions were classified as CVS (“50% rule”), as MS-mimic otherwise.
Lesion-level performance was evaluated by specificity, sensitivity, and area under the curve (AUC) on the test set. Patient-level performance was evaluated on the validation set (28 patients) due to the small size of the test set.
Results: Compared to expert visual assessment, at lesion level the CNN reached 81% sensitivity, 71% specificity, and 79% AUC, with near-instant classification (vesselness filter 81%, 57%, 74% resp., around .5 seconds per lesion). At the patient level, the CNN had sensitivity 91%, specificity 88% (vesselness filter 50%, 82% resp.).
Conclusions: The relatively high performance and speed of the proposed automated approaches are promising and show potential for clinical applicability. The CNN performed better than the vesselness filter, without need for lesion masking. Relatively simple improvements to the CNN training and architecture should further increase its performance, which might ease translation of the CVS imaging biomarker into clinical practice.
Disclosure: Jonas Richiardi: Salary paid in part by Siemens Healthcare AG Switzerland
Pietro Maggi: Supported by the ECTRIMS Clinical Training Fellowship Program
Mário João Fartaria: Salary paid by Siemens Healthcare AG Switzerland
Francesco La Rosa: Nothing to disclose
João Jorge: Nothing to disclose
Pascal Sati: Nothing to disclose
Daniel S. Reich: No disclosures relevant to the content of this abstract
Renaud Du Pasquier: Nothing to disclose
Reto Meuli: Nothing to disclose
Meritxell Bach Cuadra: Nothing to disclose
Tobias Kober: Salary paid by Siemens Healthcare AG Switzerland