Author Of 1 Presentation
PS16.04 - RimNet: A deep 3D multimodal MRI architecture for paramagnetic rim lesions assessment in multiple sclerosis
Abstract
Background
In multiple sclerosis (MS), perilesional chronic inflammation appears on in vivo 3T susceptibility-based magnetic resonance imaging (MRI) as non-gadolinium-enhancing paramagnetic rim lesions (PRL). A higher PRL burden has been recently associated with a more aggressive disease course. The visual detection of PRL by experts is time-consuming and can be subjective.
Objectives
To develop a multimodal convolutional neural network (CNN) capable of automatically detecting PRL on 3D-T2*w-EPI unwrapped phase and 3D-T2w-FLAIR images.
Methods
124 MS cases (87 relapsing remitting MS, 16 primary progressive MS and 21 secondary progressive MS) underwent 3T MRI (MAGNETOM Prisma and MAGNETOM Skyra, Siemens Healthcare, Erlangen, Germany). Two neurologists visually inspected FLAIR magnitude and EPI phase images and annotated 462 PRL. 4857 lesions detected by an automatic segmentation (La Rosa et al. 2019) without overlap with PRL were considered non-PRL. The prototype RimNet was built upon two single CNNs, each fed with 3D patches centered on candidate lesions in phase and FLAIR images, respectively. A two-step feature-map fusion, initially after the first convolutional block and then before the fully connected layers, enhances the extraction of low and high-level multimodal features. For comparison, two unimodal CNNs were trained with phase and FLAIR images. The areas under the ROC curve (AUC) were used for evaluation (DeLong et al. 1988). The operating point was set at a lesion-wise specificity of 0.95. The patient-wise assessment was conducted by using a clinically relevant threshold of four rim+ lesions per patient (Absinta et al. 2019).
Results
RimNet (AUC=0.943) outperformed the phase and FLAIR image unimodal networks (AUC=0.913 and 0.855, respectively, P’s <0.0001). At the operating point, RimNet showed higher lesion-wise sensitivity (70.6%) than the unimodal phase network (62.1%), but lower than the experts (77.7%). At the patient level, RimNet performed with sensitivity of 86.8% and specificity of 90.7%. Individual expert ratings yielded averaged sensitivity and specificity values of 76.3% and 99.4%, respectively.
Conclusions
The excellent performance of RimNet supports its further development as an assessment tool to automatically detect PRL in MS. Interestingly, the unimodal FLAIR network performed reasonably well despite the absence of a paramagnetic rim, suggesting that morphometric features such as volume or shape might be a distinguishable feature of PRL.
Author Of 2 Presentations
P0017 - Translating MPRAGE to MP2RAGE improves the automatic tissue and lesion segmentation in Multiple Sclerosis patients (ID 766)
Abstract
Background
Compared to the conventional magnetization-prepared rapid gradient-echo imaging (MPRAGE) MRI, magnetization prepared 2 rapid acquisition gradient echoes (MP2RAGE) (Marques, J. P., 2010) shows a higher brain tissue and lesion contrast in multiple sclerosis (MS) patients. This specialized sequence is, however, mainly limited to research settings and not widely acquired in clinical routine.
Objectives
To synthesize realistic-looking MP2RAGE images from MPRAGE acquisitions via generative adversarial network (GAN) and verify if these improve the performance of automatic MS lesion and brain tissue segmentation tools.
Methods
We propose a GAN inspired by the pix2pix framework (Isola, P., 2018) which takes as input 2D slices of 3D MPRAGE images and generates a synthetic MP2RAGE (synMP2RAGE). Differently from pix2pix, the generator of the GAN combines three loss functions: a pixel-wise L1 loss, an adversarial loss, and a perceptual loss. Our framework is trained on 12 healthy controls and 8 MS patients and tested on 36 MS patients, for which an expert manually delineated cortical and white matter lesions. Imaging was performed with a 3T MRI scanner (Siemens Healthcare, Erlangen, Germany) with a 1x1x1.2 mm resolution. Evaluation is performed with reference-based metrics and through automatic segmentation of MS lesions (La Rosa, F., 2020) and brain tissue (Avants, B.B., 2011).
Results
Considering as reference the acquired MP2RAGE, synMP2RAGE achieves a peak signal-to-noise ratio of 31.39, normalized root mean square error of 0.13, and structural similarity index of 0.98, overperforming the MPRAGE (29.49, 0.17, 0.97, respectively) for all metrics. Compared to the initial MPRAGE it also significantly improves the lesion and tissue segmentation masks in terms of the Dice coefficient and volume difference (p-values < 0.001). On the contrary, no significant differences between the real and synMP2RAGE are found in the patient-wise comparison of the lesions’ segmentation (p-values > 0.05), whereas they are significant between MPRAGE and MP2RAGE (p-value < 0.001).
Conclusions
Our proposed framework successfully translates MPRAGE to MP2RAGE, synthesizing realistic-looking images which improve the performance of automatic segmentation tools tested on MS patients. In accordance with previous claims (Finck, T., 2020), these results confirm that GANs can be helpful in the automatic analysis of MRI images.
P0545 - Automatic MS lesions segmentation using LeMan-PV as a clinical decision-support tool: a longitudinal analysis (ID 1590)
Abstract
Background
LeMan-PV is a prototype that performs cross-sectional and longitudinal detection of Multiple Sclerosis (MS) lesions, which has been validated on conventional (cMRI) and advanced magnetic resonance imaging at 3T (Fartaria et al. 2019). Since this software provides a report that is available shortly after image acquisition, it may be ideal as clinical decision-support tool.
Objectives
To assess LeMan-PV as clinical decision-support tool in a monocentric real-world cMRI dataset from the Swiss Multiple Sclerosis Cohort.
Methods
262 MS patients underwent cMRI at Basel University Hospital in a mean of 3.5 follow-up sessions, with an average of 399 days between two consecutive sessions. cMRI sequences were acquired at 1.5T and 3T in 725 and 195 sessions, respectively. Cross-sectional and longitudinal MS lesions segmentation (i.e. identification of new and enlarging lesions - NLs, ELs) was performed using the LeMAN-PV prototype software. An expert neuroradiologist performed a radiological reading of the number of NLs and ELs in the most recent acquisition by comparing it to the previous one (ground truth, GT), considering only lesions with a diameter larger than 3 mm. The minimum volume thresholds to identify an NL and an EL were chosen by minimizing the patient-wise error between the automated count and the expert ground truth. Two scenarios were evaluated by first assuming disease activity if one or more EL were present, and second by considering activity if NL were present in the new acquisition.
Results
The volume thresholds chosen were 11 and 12 mm3 for ELs and NLs, respectively. For those, LeMan-PV detected 11% more of both ELs and NLs than the neuroradiologist. In the patient-wise evaluation of cases with both sessions acquired at 1.5T (70%), LeMan-PV showed sensitivities of 93% and 78% and specificities of 62% and 43% when evaluating ELs and NLs. For the 3T pairs of sessions (8%), values were 68% and 72% for ELs and 73% and 68% for NLs. Finally, for cases with a first acquisition at 1.5T and a second at 3T (22%), values were 76% and 73% for ELs and 71% and 65% for NLs.
Conclusions
The count of new and enlarging MS lesions using LeMan-PV were close to the one performed by an expert neuroradiologist; the software performed better when assessing disease activity via detection of enlarging lesions rather than by identifying new lesions. More 3T data is being currently collected at 3T to provide a size-matched inter-scanner comparison.