Triplanar U-Net with lesion-wise voting for the segmentation of new lesions on longitudinal MRI studies

Abstract

We present a deep learning method for the segmentation of new lesions in longitudinal FLAIR MRI sequences acquired at two different time points. In our approach, the 3D volumes are processed slice-wise across the coronal, axial, and sagittal planes and the predictions from the three orientations are merged using an optimized voting strategy. Our method achieved best F1 score (0.541) among all participating methods in the MICCAI 2021 challenge Multiple sclerosis new lesions segmentation (MSSEG-2). Moreover, we show that our method is on par with the challenge's expert neuroradiologists: on an unbiased ground truth, our method achieves results comparable to those of the four experts in terms of detection (F1 score) and segmentation accuracy (Dice score).

A generalizable deep voxel-guided morphometry algorithm for the detection of subtle lesion dynamics in multiple sclerosis

“…Moreover, there exist multiple automated new lesion segmentation algorithms based on deep learning (Andresen et al, 2022 ; Ashtari et al, 2022 ; Basaran et al, 2022 ; Hitziger et al, 2022 ; Kamraoui et al, 2022 ; Sarica and Seker, 2022 ; Schmidt-Mengin et al, 2022 ; Commowick et al, 2023 ). These methodologies primarily leverage the MSSEG-2 dataset (Commowick et al, 2021 ), encompassing FLAIR images from baseline and follow-up visits for each patient, either with or without the integration of synthetic data.…”

Influence of Residual Quadrupolar Interaction on Quantitative Sodium Brain Magnetic Resonance Imaging of Patients With Multiple Sclerosis

“…The binary masks were generated from these tissue probability maps by comparing the tissue contributions for each voxel and choosing the maximum value as corresponding tissue. The lesion mask was created using the FLAIR image and the AI-based segmentation software mdbrain (Mediaire GmbH, Berlin, Germany) 32 . All voxels characterized as lesions by mdbrain were excluded from the NAWM, NAGM, and CSF masks.…”

“…The lesion mask was created using the FLAIR image and the AI-based segmentation software mdbrain (Mediaire GmbH, Berlin, Germany). 32 All voxels characterized as lesions by mdbrain were excluded from the NAWM, NAGM, and CSF masks. Lesion masks were created not only for MS patients but also for HC to account for lesion-like FLAIR hyperintensities arising from aging or other processes.…”

Toward more accurate diagnosis of multiple sclerosis: Automated lesion segmentation in brain magnetic resonance image using modified U‐Net model

“…Prediction using all directions may increase the probability of detecting lesions, as each axis will expose different volumes of lesions. 52,53 This scheme is inspired by Hitziger et al 54 and Aslani et al 7 However, it is different in terms of how they merge predictions from all axes. The first used the average of the predicted probabilities and then compared it with three different fusion techniques (union, majority, and intersection), while the latter used the major voting method.…”

“…The CF strategy comes as an alternative to two simple operations (i.e., union and intersect) as used by past studies. 7,29,34,54 This method is inspired by the calculation of lesion-wise metrics performance of two raters, as explained by Carass et al 55 To better understand this technique, a fusion of two simple 1D binary predictions is presented in Figure 4. Thus, instead of finding similarities between two predictions at the pixel level, the technique focuses on the agreement of lesions' region location.…”