Keywords:
Artificial Intelligence, Neuroradiology brain, MR, Computer Applications-General, Segmentation, Dementia, Image verification, Tissue characterisation
Authors:
M. I. Ferraz Meyer, S. van Eyndhoven, S. Vercruyssen, V. Terzopoulos, P. Frenyo, A. Brys, D. Smeets, D. M. Sima
DOI:
10.26044/ecr2023/C-17037
Methods and materials
The icobrain software has recently been equipped with a deep learning (DL)-based brain segmentation algorithm. In an effort to achieve results which are consistent across MR images acquired in different settings (e.g., scanner or centre) the dataset used for training contained a lot of variability both at the population level and in terms of scanner- and center- or acquisition-specific factors (i.e., age ([16, 81] years), sex (M/F ∼ 33%/67%), slice thickness in T1 ([0.4, 1.5] mm), magnetic field strength (1.5T/3T ∼ 43%/57%), scanner manufacturer (Philips, GE, Siemens and Hitachi), scanner model and acquisition sequence). Additionally, a specific intensity-based augmentation strategy was used, as presented in [4].
- We evaluate the reproducibility of the method on a test-retest dataset consisting of T1 and FLAIR images from 10 multiple sclerosis (MS) patients, who were scanned twice (with repositioning) in three scanners from different vendors. Further details about the data can be found in [5].
- We estimate the difference in computed volume for each of the evaluated structures between the same scanner repetitions (intra-scanner differences) and between the repetitions performed in different scanners (inter-scanner differences).
- The performance is then compared against the previous version of the software (expectation-maximisation algorithm, v5.9) and a state-of-the-art DL approach (FastSurfer) [6].