Participants
The neurologist made a diagnosis of temporal lobe epilepsy (TLE) through clinical exams and EEG monitoring. All patients have an MRI exam. Structural MR images (n = 118) were obtained from individuals aged 13–59 years at the Clinic of Neurology, University Clinical Center of Serbia, Belgrade. The participants included 54 men (45.8%) and 64 women (54.2%). They encompassed individuals with temporal lobe epilepsy (TLE) (n = 80) and healthy participants (n = 38). Individuals with TLE were divided into two groups, the TLE-R group (n = 28) and the TLE-L group (n = 52). The TLE-R group included 14 men (50%; mean age, 31.3 ± 7.6 years) and 14 women (50%; mean age, 30.4 ± 7.9 years). The TLE-L group included 29 men (55.8%; mean age, 35.6 ± 9.1 years) and 23 women (44.2%; mean age, 34.3 ± 10.2 years). Healthy participants included 17 men (44.7%; mean age, 27.3 ± 6.2 years) and 21 women (55.3%; mean age 35.7 ± 9.9 years.
Mr Imaging Data
The T1-weighted (T1w) MR imaging sequence was used for anatomical characterization. High-resolution structural scans were acquired in the sagittal plane at 1.5 T (Philips Achieva, Amsterdam, Netherlands), 1.5 T (Siemens, Avanto/Aera, Erlangen, Germany), and 3 T (Siemens Skyra, Erlangen, Germany). The imaging parameters were as follows: TR = 25ms, flip angle = 8°, acquisition matrix of 256×232×180 (for 1.5 T Philips); TR = 1700ms, flip angle = 4°, acquisition matrix of 240×256×180 (for 1.5 T Siemens); and TR = 1700ms, flip angle = 5°, acquisition matrix of 240×256×180 (for 3 T Siemens). The voxel size for all three MRI machines was 1×1×1 mm3.
Segmentation Methods
QuickNAT segmentation
Quick segmentation (QuickNAT) software is based on a deep fully convolutional neural network (F-CNN). Its postprocessing times run in seconds and use three two-dimensional F-CNNs, where each operates in a separate plane (transversal, sagittal, and coronal). The first step of making QuickNAT was to apply the existing software FreeSurfer [17] to segment scans without annotations. After that, the software was used to continue training the previous network with smaller manually annotated data to achieve high segmentation accuracy. QuickNAT was the first to utilize the publicly available software (FreeSurfer) to segment the data and train the network. The authors reported high test-retest accuracy, which makes it suitable for longitudinal studies. The final segmentation and labels correspond to 27 brain structures (cortical and subcortical). The code and trained model are available as extensions of MatConvNet [18] at https://github.com/abhi4ssj/QuickNatv2.
Fsl-first Segmentation
A commonly used, automated, model-based approach is the FMRIB Integrated Registration and Segmentation Tool (FIRST), which is provided as a part of the FSL software library [15].
The construction of the model was based on a manually labeled dataset [19]. It first registers the images to MNI152 space by performing affine registration. FSL-FIRST [20] is one of the most commonly used algorithms for the segmentation of subcortical structures and has been validated by the scientific community [21, 22]. It uses a model-based single atlas that runs approximately about 15 minutes. Before starting segmentation, we used the “fslreorient2std” script to ensure the adequate orientation of the images. The “run_ first_all” script was applied to automatically segment subcortical structures into 15 different labels. The fully automated segmentation of the volumes is available at http://fsl.fmrib.ox.ac.uk/. We used FSL-FIRST v.6.0.3.
Quality Control Of Segmentation
Segmentations with gross errors were excluded from further investigation. These errors typically happened due to a major failure of the atlas to target image registration. Images requiring exclusion were images where the segmentation region of interest (ROI) did not overlap with anatomical brain boundaries. All segmentation images have a visual inspection.
Determination Of Common Roi Labels
A direct comparison of an ROI is not possible because the references of labels and the delineation of boundaries of anatomical regions are different between QuickNAT [17] and FSL (Harvard-Oxford atlas). Therefore, we used a FreeSurfer space for overlapping the results of the segmentation of these two different methods. The set of ROIs included the following subcortical structures: thalamus (L + R), caudate (L + R), putamen (L + R), pallidum (L + R), and amygdala (L + R). Matched ROIs were inspected by one manual rater (Z.J.) based on the ROI’s names and spatial correspondence between them. A visualization of the ROIs and list of matched ROIs at QuickNAT and FSL-FIRST is presented in Fig. 1 and Fig. 1A.
Statistical analysis
Statistical analyses were performed in Rstudio software (version 1.4.1106). The percentage volume difference (PVD) was calculated by using the following formula:
$$PVD=100 x \frac{|V\left(A\right)-V\left(B\right)|}{\frac{V\left(A\right)+V\left(B\right)}{2}}$$
where V(A) and V(B) were volumes obtained by using FSL and QuickNAT, respectively.
Comparison between volumes obtained by different techniques (parameters – brain segmentations) on the same group of subjects (separately TLE-R, TLE-L, and control) was done by using paired t test or Wilcoxon test. t test was used when the normality of the distribution is satisfied, and Wilcoxon test when is not. Kruskall-Wallis test was used to determine whether or not a statistically significant difference exists between TLE-R, TLE-L, and the control group of subjects. For pairwise comparisons between each independent group, if the results of Kruskall-Wallis test were statistically significant, was used Dunn’s test. The Bonferroni correction was used for Dunn’s test, so the p-values considered significant only if it was smaller than 0.05/3 = 0.017.
Bonferroni correction was used because we had multiple comparations. In our study, we had a 10 comparison and needed to divide with a limit value (p < 0.05). That is a reason why p < 0.005 was considered significant. Intraclass correlation coefficients (ICCs) were calculated to describe the consistency between the volumes obtained from FSL and QuickNAT. An ICC calculation two-way mixed model was selected. An ICC value of 1 indicated perfect reproducibility between the two methods, while a value of 0 suggested reproducibility lower than expected based on chance alone.
The Dice coefficients are a statistical test used for the validation of imaging segmentation algorithms. We used FSL-FIRST as a gold standard because it has been validated in the scientific field. The Dice coefficient represents the spatial overlap index. The value ranges from 0, which indicates no spatial overlap between the two segmentation methods, to 1, which indicates complete overlap. The Dice coefficients (D) were calculated by using the following formula:
$$D=2*\frac{|X\bigcap Y|}{\left|X\right|+\left|Y\right|}$$
where X and Y were segmentations obtained by using FSL and QuickNAT, respectively; |X| and |Y| means the number of elements in that set (X or Y); ⋂ represented the intersection of the two segmentation methods; and means of the elements that are common in both methods.