This study included 104 Cuban patients aged between 60 and 87 years (73.77 ± 6.74 years) from the Hospital General Docente Iván Portuondo in Artemisa Province and the Instituto Nacional de Neurología y Neurocirugía in Havana. Participants had previously been diagnosed with possible mild or moderate AD based on cognitive tests performed according to the criteria of the National Institute on Aging-Alzheimer's Association (NIA-AA) and the mild-to-moderate impairment criteria of the Global Deficit Score, also supported by the clinical and neuropsychological diagnosis.22 All participants signed informed consent forms and the study was approved by the research ethics committee of the participating institutions. For comparison, we used MRI scans of 272 healthy controls (mean age, 73.15 ± 7.53 years) from the free-access Cambridge Centre for Ageing Neuroscience (Cam-CAN) database (https://camcan-archive.mrc-cbu.cam.ac.uk/dataaccess/).

We used the pulse sequences recommended by the Alzheimer's Disease Neuroimaging Initiative (ADNI),23 adapted to the Siemens Magnetom Allegra 3.0 T scanner installed at the Centro de Neurociencias de Cuba (CNEURO). Total study time per patient ranged from 38 to 40 min; patients were instructed to remain still, awake, with their eyes closed, in the supine position. We acquired 3D sagittal sequences with axial and coronal reconstruction. In this study, we used the T1-3D MP-RAGE sequence (voxel size [mm3]: 0.8 × 0.8 × 1; time [ms]: TE = 2.6; TR = 2000; TI = 900; duration [min]: 9:19) for manual and automatic segmentation of the hippocampus.

There are multiple manual segmentation protocols,24,25 whose differences and similarities are mainly based on the definition of anatomical limits. In an international effort to standardise existing protocols in 2014, researchers from the European Alzheimer’s Disease Consortium (EADC) and the ADNI created the EADC-ADNI Harmonised Protocol for Hippocampal Segmentation (HarP, appendix II) (www.hippocampal-protocol.net).20 We used this protocol because it enables the training of new master tracers26 and improvement of automatic algorithms27; international experience has also demonstrated high reproducibility of measurements.15 To identify the side of the hippocampus, MRI scans were displayed according to the radiological convention.

Manual calculation of hippocampal volume was performed using the Amira 5.3.3 program (Visage Imaging, Berlin, Germany; http://www.visageimaging.com) running on a computer with an 8th generation Intel® Core i5 processor, Windows 10 operating system, and 16 GB RAM. We selected this program based on our working group’s previous experience with this tool.28 For the automatic segmentation of the hippocampus of healthy controls and patients with AD, we used the FreeSurfer 6.0 software, which is available free of cost for Linux and macOS (http://surfer.nmr.mgh.harvard.edu/). This software performs segmentation of brain structures, including the hippocampus, using probability maps.29 Image processing was performed with a high performance computer (HPC) system with a peak performance of 4 Tflops. The HPC system includes 5 computing nodes with 2 Intel Xeon CPUs with 24 cores each, for a total of 48 cores per node, with 48 GB of RAM and two 250-GB HDDs each. The system has a 24-TB network access storage system, which stores all the processed data.

All images were visually inspected before processing, in accordance with quality control procedures previously described in the literature.30 Digital Imaging and Communication in Medicine (DICOM) images were converted to the Neuroimaging Informatics Technology Initiative (NIfTI) format using the multiplatform MRIcroGL software (https://www.nitrc.org/plugins/mwiki/index.php/mricrogl:MainPage), and subsequently processed with FreeSurfer 6.0, using the cross-sectional pipeline recon-all command. Reports of segmentation findings in the literature have been very heterogeneous, as some authors do not normalise for head size. Therefore, this study reports both measurements. The asymmetry index (AI) was calculated based on the differences between absolute values for the left and right hippocampal volumes normalised using the following formula:

Statistical analysis was performed using the IBM SPSS Statistics software, version 26.0.

Comparison between methods:

• •

  Normality of the measurement results was tested with the Kolmogorov–Smirnov test.

• •

  Reproducibility of the manual measurements was studied with the intraclass correlation coefficient (ICC) with a 95% confidence interval (CI) (Model: Alpha, Two-Way Mixed; type: absolute-agreement).31

• •

  General concordance between measurements taken with manual and automatic methods was assessed with the Bland–Altman plot with a 95% CI.

• •

  The ICC with a 95% CI was used to assess measurement agreement of both methods.

• •

  We used the 2-tailed unpaired t test (P < .05) to identify significant differences between measurements for both hippocampi.

Parametric statistics (non-normalised hippocampal volume):

• •

  We used the 2-tailed unpaired t test (P < .05) to compare both hippocampi for the same sex and between all individuals.

• •

  We used the one-tailed unpaired t test (P < .05) to compare the left and right hippocampi.

• •

  We used analysis of variance (ANOVA, P < .05) to study significant differences in hippocampal volume per age group, as processed with the automatic method.

Non-parametric statistics (normalised hippocampal volume):

• •

  To analyse the different age groups whose data were processed using the automatic method, we used the Mann–Whitney U test or Kruskal–Wallis test (P < .05).

To assess intra-observer error, the manual measurement method was applied to a single individual 10 different times, with no time limit. The reliability of the procedure was assessed using the ICC with a 95% CI.32

After analysing the usefulness of the Amira software for manual segmentation of the hippocampus, we determined which modules to use (Fig. 1). To load DICOM images corresponding to the T1-3D MP-RAGE sequence, the software uses the option “Load File” and stores the object as “MPRAGE-SAG-COR” together with the reconstruction of the coronal and axial planes (module marked as A).

Fig. 1.

Pool of modules and objects used during the processing of each individual, with connecting lines that indicate the dependent processes; items are alphabetically identified according to the order of execution.

(0.09MB).

We subsequently inserted the OrthoSlice module (B) to reconstruct coronal and axial planes. Segmentation of each hippocampus was performed by creating two separate objects, one for each hippocampus (C). To do this, we used the tool “Segmentation Editor” and proceeded according to the HarP protocol. Once segmentation was completed, we generated the surfaces by adding the SurfaceGen module (D), which generates the “.surf” objects (E). These objects contain 3D information for each hippocampus, which may be displayed in the SurfaceView module (F). Lastly, after adding the MaterialStatistics module (G), we performed the volumetric calculation, resulting in generation of the “.MaterialStatistics” (H) object, which contains the report on hippocampal volume (Table 1).

Using the above-mentioned modules, we obtained the reconstruction, visualisation, and quantification of the left and right hippocampal volume of each processed case (Fig. 2).

Fig. 2.

Three-dimensional reconstruction of the manual segmentation of the hippocampus, performed using the Amira software, from three study participants. The left hippocampus is shown in yellow and the right in magenta. Left Hip Vol: volume of the left hippocampus; Right Hip Vol: volume of the right hippocampus.

(0.19MB).

To assess the reliability of the manual segmentation method, the procedure was performed 10 times for a single study participant (Table 2). We obtained a coefficient of variation of 2.50% for the calculation of left hippocampal volume and of 2.65% for the right.

To be considered a master tracer in the HarP project, researchers must obtain an ICC ≥ 0.90 in the calculation of hippocampal volume.26 This coefficient was calculated for each hippocampus, with the 5 first and 5 last measurements being compared. The ICC was calculated using both types of the two-way random model: consistency and absolute agreement. As a result of the analysis, we obtained ICC values of 0.95 in the consistency test and 0.93 for absolute agreement for the left hippocampus. For the right hippocampus, we obtained an ICC of 0.90 in the consistency test and 0.90 for absolute agreement. These values are acceptable according to the HarP criteria; therefore, the procedure is reliable for clinical use.

Processing with FreeSurfer was performed using the recon-all command for cross-sectional analysis: recon − all − s < output folder name >−i < subjID.nii >−all − 3T − qcache.

We used the -3T flag to correct image intensity,33 making it more appropriate for 3T MRI.34 We also used the FreeSurfer atlas based on 3T for the alignment of Talairach coordinates. The volume of each region of interest (ROI) from the FreeSurfer atlas was directly obtained from the “aseg.stats” output files. FreeSurfer segmentation outputs were visually inspected to detect serious errors. In the event of serious errors, or if FreeSurfer crashed without providing results and the issue could not be resolved after repeated attempts, analyses of these images were omitted.

It is important to highlight that in the FreeSurfer software, it is sometimes necessary to manually correct results using Freesurfer’s own viewer: Freeview. In this case, after the necessary manual corrections were made to the “aseg.presurf.mgz” file (when version 6.0 is used), the following command must be executed in the console: recon − all − autorecon2 − noaseg − autorecon3 − subjid < Subj_ID >.

This command recalculates the volumes with the new corrections and generates a new “aseg.stat” file. No manual corrections were made in this study because our objective was to compare the results of both procedures. We should underscore that the FreeSurfer development team advise against correcting minor errors. Minor errors are defined as those that only appear in one slice; therefore, correcting this single error would not significantly modify results. The program presents good reproducibility, and manipulating the images excessively may bias results.35

In our study, we only used the T1-3D MP-RAGE sequence in processing; however, it may also be used with high-resolution T2-weighted sequences. This is known as multimodal processing.36

Fig. 3 shows the behaviour of volume values as a result of manual and automatic segmentation of 52 study participants. We observed a trend towards higher values with the automatic method, which is more evident in the right hippocampus (Fig. 3B). This finding has previously been reported in the literature.19 After visually inspecting the segmentation of patient 40, we confirmed the overestimation of volume that FreeSurfer tends to make (Fig. 4). However, comparison of the measurement results from both methods with the t test detected no significant differences for the left hippocampus (P = .667) or the right hippocampus (P = .105). Although there were no significant differences between methods, we observed incongruences in results for some patients, such as patients 35 and 40 (Fig. 3A). This confirms the need to supervise the process.

Fig. 3.

Left and right hippocampal volumes obtained using the manual and automatic segmentation methods. The black circles indicate two atypical values from the automatic segmentation as compared to manual segmentation.

(0.16MB).

Fig. 4.

Comparison of results obtained with Amira and FreeSurfer for the segmentation of both hippocampi of a 72-year-old study participant. The red circles indicate areas outside the hippocampus that were incorrectly included by FreeSurfer. This is one of the reasons for the overestimation of volumes.19

(0.23MB).

The Bland–Altman plot showed good concordance of the results for both hippocampi, as the differences identified were close to the mean, within the CI (95%), and did not show a relational trend (Fig. 5). ICC calculation for the left and right hippocampi displayed excellent reliability, with a value of 0.95 for both hippocampi. The lower confidence limit of the ICC of the right hippocampus was 0.659 (considered moderate according to Shrout et al.31), which contrasts with the excellent agreement reported in the left hippocampus. This may be due to the fact that the right hippocampus consistently presents more differences between the results of automatic segmentation than of manual segmentation. However, values were similar in most cases. Therefore, the automatic method is appropriate for processing large numbers of individuals, and must be supervised. This requires quality control prior to image processing,30 and quality control of segmentation.

Fig. 5.

Bland–Altman plot analysing agreement between the manual and the automatic methods for calculating the volumes of the left and right hippocampi. The difference (central bold line) was 41.44 mm3 (SD = 213.6) for the left hippocampus and 154 mm3 (SD = 1579) for the right hippocampus. Limits of agreement are expressed as mean ± 1.96 SD and shown as red dotted lines. SD: standard deviation.

(0.13MB).

After analysing the reliability of automatic processing, we completed hippocampal volume quantification for all 104 participants (Table 3). We found no significant differences (P = .08) when comparing the volumes of both hippocampi. This contradicts previous reports in the literature, which describe greater right hippocampal volumes.37 Similarly, comparison of both hippocampi did not reveal significant differences for either sex (P = .20 for women and P = .50 for men) or between sexes (P = .34 for the left hippocampus and P = .40 for the right).

The mean values obtained are similar to those described in previous reports from Latin America.20 These, in turn, show significant differences with regard to those observed in healthy controls from Europe (Table 4).

The values obtained with automatic processing showed a trend to decreased hippocampal volume as individuals aged (Table 5). ANOVA showed significant differences between the first and the last age groups for both hippocampi (P = .0018), as shown in Table 5.

Tabla 5.

Mean right and left hippocampal volumes obtained by automatic processing of 104 individuals, by age group. Graphs show ANOVA analysis of measurements by age group for each hippocampus.

Age group (n)	Left hippocampus (mm3)		Right hippocampus (mm3)
60-69 (n = 32)	3083 ± 531.8		3105 ± 565.3
70-79 (n = 47)	2840 ± 528.0		2933 ± 513.4
80-89 (n = 25)	2580 ± 384.7		2612 ± 454.3

However, hippocampal volume is influenced by head size; therefore, measurement should be normalised with the estimated total intracranial volume (eTIV) for each individual.38,39 When comparing the normalised measurements of our study participants with those in the Cam-CAN database, we observed that our measurements were below the mean (Fig. 6A and C). Significant differences were found for both hippocampi (P = .0033 for the left and P = .0003 for the right).

Fig. 6.

(A and C) Scatter plots of the eTIV-normalised volumes of both hippocampi adjusted to age of the individuals from CNEURO (patients), compared to individuals from Cam-CAN (healthy controls). (B and D) Box-and-whisker plots of the analysis of significant differences between age groups in the CNEURO study; the Kruskal–Wallis test was used to analyse eTIV-normalised hippocampal volumes.

(0.33MB).

The Mann–Whitney U test identified no significant differences between left and right hippocampal volumes in patients with AD. The Kruskal–Wallis test only revealed significant differences between the first and last age groups for both hippocampi, showing two levels of statistical significance (Figs. 6B and D) (P = .0018 for the left, and P = .0015 for the right).

Analysis of hippocampal asymmetry values in our patient sample did not reveal significant differences between sexes (P = .117). Of the 104 patients, 54 showed right lateral asymmetry. While this was not significant, it contradicts previous reports that describe greater atrophy in the left hemisphere.16 However, the mean asymmetry values from Cam-CAN and CNEURO show statistically significant differences between both databases for all subjects and for women, as shown in Table 6. One limitation of our study is not including a control group of healthy individuals from Cuba. However, this study represents another step towards the creation of normative brain databases of the Cuban population.40