Automated detection of cortical lesions with single and multi-contrast 7T MRI
AUTHORS: La Rosa F, Beck ES, Maranzano J, Thiran JP, Granziera C, Reich DS, Sati P, Bach Cuadra M
ECTRIMS 2021 - Multiple Sclerosis Journal, 27(2_SUPPL): 458-459, October 2021
Introduction: Cortical lesions (CL) are common and often extensive in multiple sclerosis (MS) and are associated with disability (Harrison, et al., JAMA Neurol, 2015). Detection of CL is challenging with 3T MRI, but it improves with a combination of high-resolution magnetization prepared 2 rapid acquisition gradient echoes (MP2RAGE) and T2*-weighted imaging at 7T (Beck, et al., AJNR, 2018). We have recently proposed a framework for the automated detection of CL using these two image types (La Rosa, et al., MICCAI, 2020).
Objectives: First, we modified a convolutional neural network (CNN) from our previous work to improve its performance at detecting CL given three 7T MRI contrasts (T1w MP2RAGE, T2*w EPI, and T2*w GRE) as input. Second, we developed and assessed a unimodal architecture based on only MP2RAGE and compared the results to the previous CNN.
Methods: We retrospectively analysed the MRI of 60 MS patients acquired at 7T with MP2RAGE (average of 4 acquisitions), T2* GRE, and T2* EPI (resolution of each sequence: 0.5mm isometric). 2396 CL were manually annotated by two experts, who reached consensus in a joint session. Compared to our recently proposed framework (La Rosa, et al., MICCAI, 2020), three main modifications were performed. First, the ground truths were improved, adding those CNN’s false positives that were confirmed as true CL by an expert. Second, an additional resolution level in the CNN architecture was added. Third, we used the focal loss during training instead of cross-entropy, weighting the contribution of each voxel based on the CNN’s segmentation error. An architecture with the three MRI contrasts as input and another one with only the MP2RAGE were trained and evaluated in a 6-fold cross-validation.
Results: The three-contrast CNN achieved a lesion-wise true positive rate (LTPR) of 73%, lesion-wise false positive rate (LFPR) of 38%, and Dice coefficient (DSC) of 0.47. The CNN that used only MP2RAGE achieved a LTPR of 72%, LFPR of 38%, and DSC of 0.44. There were no method-related differences in these measures on a patient level (Wilcoxon signed-rank test). However, both CNNs significantly (p<0.05) outperformed our previous results on the same dataset (LTPR=67%, LFPR=42%, DSC=0.39).
Conclusions: With a few targeted modifications, we have boosted the performance of our previously proposed framework for CL detection at 7T (LTPR of 73% vs 67%). Use of T2*w images did not significantly improve performance vs using MP2RAGE images alone.