Data acquisition was performed following the Good Clinical Practice guidelines and approved by an appropriate Ethical Committee (2022-A01030-43). Written informed consent was obtained from each participant. Forty-seven male volunteers participated in this study, including 17 freedivers (FD) and 20 controls (NC) aged 27–55 years and recruited between 2022 and 2024 in the pools of Ile de France region (Paris, Yvelines, Hauts de Seine, France). A schematic of our protocol is shown in Fig. 1.

Fig. 1.

Acquisitions calendar over one sporting season. 3T anatomical MRI and cognitive tests (HippoPS task) before and after homogeneous freediving training 3 times a week, including 1 hour 30 minutes of training for each session, and for non-freediver sport controls. FD, freedivers; NC, controls; MRI, magnetic resonance imaging; PS, pattern separation.

Cerebral MRI was performed at 3T (VE11, 3T PRIMA Fit Siemens Healthineers, Erlangen, Germany) including the acquisition of T1 and high-resolution T2-weighted images of the HS. T1 parameters were Repetition Time (TR) = 2300 ms, Echo Time (TE) = 3.05 ms, 0.9 $\times$ 0.9 $\times$ 0.9 mm resolution, 176 slices, Time = 4.44^′. T2-weighted MRI parameters were TR = 9600 ms, TE = 83 ms, Field of View (FOV) 192 mm, 0.5 $\times$ 0.5 $\times$ 1 mm resolution, 60 slices, Time = 6.46^′.

The HS were automatically segmented using the Hippocampal Segmentation Factory (HSF©) software, a segmentation tool using deep learning. Starting with a preprocessing step where MRI scans are prepared for segmentation, the HSF© pipeline entails a standalone tool, ROILoc, to extract the left and right hippocampi from the raw T2 MRI. The hippocampi are cropped with a safety margin and normalized (Z-normalization). This preprocessing step also registers MRI data to a standardized template, enabling consistent localization of hippocampal regions. Additionally, the system incorporates test-time augmentation, generating multiple versions of each HS through flips, elastic deformations, and affine transformations, which are aggregated for final segmentation.

HSF© produces voxel-wise segmentation and uncertainty maps, assigning confidence levels to each voxel. It segments the major HS, including the DG, essential for PS processes; the CA1, CA2, CA3, and the subiculum, crucial regions in episodic memory.

This approach addresses challenges in segmenting every type of HS by leveraging high-quality datasets and an efficient training process. HSF© achieves results closer to manual segmentations than existing tools, ensuring accuracy and reproducibility while significantly reducing processing time. More details about this tool can be found at https://hsf.readthedocs.io/en/latest/ [27].

Episodic memory was assessed using the HippoPS task [28] (Fig. 2), a validated ecological tool specifically designed to study PS. This task is particularly suited for evaluating episodic memory, as it provides a more precise assessment of its defining characteristics compared to methods commonly described in literature.

Fig. 2.

Encoding and recollection phase of the HippoPS task. The encoding phase involves presenting images of a location paired with a gesture and a temporal context (morning, noon, or afternoon). The recognition phase takes place 15 minutes later and consists of 28 questions: is the item (i) identical? (ii) similar? (iii) new?

The experiment consists of two parts: the encoding phase and, after 15 minutes delay, the recognition phase. The experiment begins with an instruction: “Lisa is a foreign tourist coming to Paris for the first time. She wants to show you around and associate a different gesture for each place she has visited. Watch the entire video carefully so that you can discuss it with her afterward. Do you recognize these places? Have you seen these gestures before?”.

In the encoding phase, participants watch a series of items presented on video, each comprising an image of a location associated with a gesture and a temporal context (morning, noon, afternoon). After 15 minutes, the recognition phase begins, with an instruction as a transition, asking the participant to distinguish photos, with the question: (i) is it an event with the same gesture and the same location (‘identical’), or (ii) is it an event with a different location but the same gesture or an event with the same location but a different gesture (‘similar location’ or ‘similar gesture’), or (iii) is it an event with a different gesture and a different location (‘new’)?

Four examples are provided to ensure that the instructions are clearly understood. 28 items were randomly presented, 12 of which were ‘identical’ events, 12 ‘similar’ events (i.e., 6 ‘similar location’ and 6 ‘similar gesture’), and 4 ‘new’ events. The results are calculated in percentage for (i) identical, (ii) similar, and (iii) new items.

All statistical analyses were conducted using R Statistical Software 4.4.2 (https://cran.r-project.org/src/base/R-4/) and JASP Team (2024 0.19.1, https://jasp-stats.org/previous-versions/), with significance set at a 0.05 threshold. The volumes of the whole hippocampus and each HS were analyzed separately. Due to deviations from normality, a Mann-Whitney test was employed to NC and FD at F_t0 (pre-training), and NC and FD at F_t1 (post-training). Within the FD group, comparisons between F_t0 and F_t1 were performed using a Wilcoxon signed-rank test. Equality of variances was assessed using Levene’s test, and an False Discovery Rate (FDR) correction was applied to account for multiple volume comparisons. Episodic memory (EM) scores were analyzed using a two-way repeated measures ANOVA. The factors included Group effect (FD vs. NC), Item effect (3 scores: (i) identical, (ii) similar, and (iii) new items), and Training effect for FD (FD_t0 vs. FD_t1). Between-group contrasts (FD vs. NC) and within-group contrasts (FD_t0 vs. FD_t1) were assessed. Levene’s test confirmed the equality of variances. However, Mauchly’s test revealed a violation of sphericity; therefore, degrees of freedom were corrected using Greenhouse-Geisser estimates of sphericity where necessary.

No significant differences were observed between the left and right hippocampal volumes; therefore, the two hippocampi were pooled for each subject (FD_t0 p = 0.455; FD_t1 p = 0.567; NC p = 0.250). At the beginning of training (t₀), the FD and NC groups showed comparable hippocampal anatomy, regardless of the HS volume considered (NC (M = 3.69, SD = 0.37) $\approx$ FD_t0 (M = 3.67, SD = 0.31), p = 0.552, d = 0.12). This was a crucial prerequisite for the study, ensuring the validity of the comparison of anatomical changes over the seven months of training.

At the end of the training season (t₁), no significant differences were observed either within the freediver group or between freedivers and controls, whether in specific HS regions or the whole hippocampus (FD_t0 (M = 3.67, SD = 0.31) $\approx$ FD_t1 (M = 3.62, SD = 0.36), p = 0.691, d = 0.13; NC (M = 3.69, SD = 0.37) $\approx$ FD_t1 (M = 3.62, SD = 0.36), p = 0.532, d = 0.12). These results indicate that hippocampal volumes remained stable despite repeated voluntary exposure to hypoxia during training (Figs. 3,4).

Fig. 3.

Automatic HS segmentation of an NC (A–C) and FD (D–F) obtained with HSF©: dentate gyrus, CA1, CA2, CA3 regions and subiculum. (A,D) Coronal MRI slices at MNI coordinates (B,E) Zoom of a coronal MRI slice at MNI coordinates (x = 244, y = 193, z = 35) (C,F) 3D volume rendering. HS, hippocampal subfields; HSF©, Hippocampal Segmentation Factory; DG, dentate gyrus; CA, Cornu Ammonis; Sub, subiculum; MNI, Montreal Neurological Institute.

Fig. 4.

HS volumes in cm³ in the NC group before training and the FD group before and after training. p-values are the Mann-Whitney test and the Wilcoxon signed-rank test with FDR correction. Mean and standard deviation are indicated. FDR, False Discovery Rate.

In addition, there was no change in the variability whether examining specific HS volumes or the whole hippocampus before (t₀) and after (t₁) training (Fig. 4).

The Group effect was not significant before (t₀) or after training (t₁), indicating comparable scores between FD and NC (p = 0.397, $\eta$² = 0.01). The Item effect was significant (p 0.001, $\eta$² = 0.38), regardless of the group or training ((i) identical items M = 71.79, SD = 22; (ii) similar items M = 48.92, SD = 19 and (iii) new items M = 74.22, SD = 25). No significant interactions were observed.

The Training effect was significant (p = 0.002, $\eta$² = 0.11), with PS memory scores for each item being significantly higher at t₁ (M = 71.24, SD = 21.73) than at t₀ (M = 62.90, SD = 24.19). Further analysis (post-hoc tests) revealed that, for both groups before and after training, memory scores were lower for (ii) similar items (p 0.001), and higher for (i) identical items than (ii) similar items (p 0.001) (Fig. 5).

Fig. 5.

Correct answers in % in the PS task before and after training, for items (i) identical, (ii) similar, (iii) new in both groups. *, significant differences between t₀ and t₁.

Unlike studies reporting hippocampal atrophy in response to pathological or environmental hypoxia, such as COPD or high-altitude exposure [13, 14, 16, 18, 19], our results suggest that hippocampal anatomy remains stable in healthy individuals following repeated voluntary hypoxia during freediving training. This highlights the resilience of the hippocampus under controlled hypoxic conditions, warranting further confirmation in larger samples.

Previous studies on accidental hypoxia consistently highlight the vulnerability of the hippocampus to hypoxia, often leading to structural atrophy associated with memory impairments [15, 17]. While many studies have focused on total hippocampal volumes or specific HS such as CA1 or the DG [4, 33], few studies have systematically examined all HS using advanced segmentation methods. Recent improvements in segmentation techniques, such as HSF©, allow detailed analysis of HS, yet their application in hypoxia studies remains limited. By employing such methods, we ensured a comprehensive examination of CA1, CA2, CA3, DG, and subiculum. Despite their sensitivity to oxygen deprivation [13, 19], no significant volume changes were detected across the freediving training period. This stability may reflect limitations in current segmentation approaches, which typically analyze the hippocampal head, body, and tail as separate regions. Further studies should consider more detailed analyses along the anterior-posterior axis or employ a gradient-based approach to better capture subtle adaptations. Furthermore, functional or microstructural changes may precede detectable volumetric changes, necessitating advanced imaging techniques for identification.

Our study also enhances the understanding of episodic memory performance under hypoxic conditions, particularly regarding PS, a process critical for distinguishing overlapping memories. The PS task assessed the hippocampus’s ability to encode unique representations for each event and tested memory consolidation by evaluating participants’ ability to distinguish between identical, similar, and new items after a 15-minute delay. Additionally, the task included associative learning components, linking actions to specific locations, and temporal contexts. Designed with ecologically valid stimuli, this task offers greater relevance to real-life memory functions compared to abstract or artificial cognitive assessments [28, 34].

We observed no significant differences in memory performance between freedivers and controls, either before or after training. However, both groups improved their PS performance post-training, consistent with the established cognitive benefits of regular physical exercise [23]. In addition, regardless of training, both groups were less adept at identifying similar items and better at recognizing identical ones. This is understandable, as distinguishing between similar items is a more complex task than recognizing a previously seen item. Thus, freedivers and controls adopt the same approach to the task. These findings contrast with studies reporting hypoxia-induced memory deficits in clinical populations, such as those with OSA [16] or COPD [14], where PS impairments were linked to hippocampal damage. Unlike these pathological conditions, freedivers who are exposed to controlled, intermittent hypoxia, preserved their memory performance. This aligns with prior studies highlighting the neuroprotective effects of physical exercise and moderate hypoxia [22, 26].

Our results are consistent with the literature on episodic memory in freedivers, involving similar cohorts in terms of the number of participants, freediving experience, age, and years of education. These studies have reported no memory deficits following hypoxemia exposure in freedivers. For instance, research using a 4-item memory task after at least 12 hours without freediving found that all divers performed within one standard deviation of published norms, suggesting no significant impairment despite years of hypoxemia exposure [21]. More recent work investigating PS performance using the Mnemonic Similarity Task (MST) found no decline in memory performance when participants completed the task 20 minutes before and 10 minutes after a freediving session [20], and testing divers 30 minutes post-exercise also reported no significant differences between freedivers and controls [35]. It is important to note that while these studies focused on acute effects, based on data collected during and immediately after performance, our study is the first to provide longitudinal data on hippocampal volume and episodic memory in the context of repeated voluntary hypoxia, allowing us to isolate the effects of freediving training on hippocampal adaptability and cognitive function.

Hippocampal plasticity is modulated by various factors, suggesting an adaptive mechanism that responds to environmental changes. Physical exercise, an intrinsic component of freediving training, plays a crucial role in hippocampal adaptability and episodic memory, promoting neurogenesis through different mechanisms [36, 37, 38]. The combination of hypoxia and physical exercise, as seen in freediving, may offer unique benefits for episodic memory, that exceed those of either factor alone. Moderate hypoxia during physical exercise has been shown to enhance episodic memory through mechanisms such as increased synaptic plasticity and neurotransmitter modulation [24]. However, the extent of hypoxia is critical; while moderate levels improve memory and neuroplasticity, severe hypoxia may impair cognitive functions [25]. The neuroadaptive benefits of freediving, who experience controlled and intermittent hypoxia, likely result from the interplay of these factors, optimizing hippocampal function while avoiding the detrimental effects associated with pathological or accidental hypoxia [39, 40].

While the frequency and duration of apnea episodes in freedivers may resemble those observed in pathological conditions, some variations exist. Both freediving and involuntary hypoxia (e.g., OSA or COPD) expose the brain to repeated oxygen deprivation, a known risk factor for hippocampal atrophy and memory deficits [13, 15]. However, the main distinction in our study lies in the voluntary nature of hypoxia during freediving, combined with regular physical exercise. This controlled exposure, paired with the physiological benefits of exercise, may explain the absence of detectable hippocampal damage or cognitive impairments in freedivers.

A recent animal study [41] suggests that controlled, moderate exposure to intermittent hypoxia can yield significant neuroprotective effects, including enhanced hippocampal angiogenesis and neurogenesis, as well as reduced brain lesion size in neonatal mice, highlighting its potential to mitigate cognitive impairments and brain injuries. Thus, carefully calibrated intermittent hypoxia could be explored as a therapeutic strategy for conditions like hypoxic-ischemic encephalopathy or other neurological disorders, opening new avenues for translational research in human medicine.

While our findings provide valuable insights, several limitations must be acknowledged. Although the HSF© software proves highly effective, outperforming other known automatic segmentation models to date (HSF’s dice coefficient compared to ASHS (p 0.001; hedge’s g = 1.64), HIPS (p 0.001; hedge’s g = 4.93), and HippUnfold (p 0.001; hedge’s g = 5.44)) [27], the MRI method and acquisition itself may lack the sensitivity needed to detect subtle changes in HS after seven months. Functional adaptations (e.g., changes in functional connectivity) might occur without being reflected in visible volumetric changes. Additionally, the stability of hippocampal volumes could reflect preserved structural homeostasis facilitated by the previously mentioned adaptations. Future studies should incorporate complementary measures such as functional connectivity (rs-fMRI), metabolic assessments (MRI spectroscopy), and cerebrovascular measurements (MRI ASL), to further explore the neurobiological mechanisms underlying adaptations to voluntary hypoxia. Another limitation is that all participants were recruited from the Île-de-France region, which may limit the generalizability of our findings to freedivers from other regions with different training regimens, and lifestyles. Additionally, including female participants in future studies could help identify potential sex differences in adaptation mechanisms and hippocampal function.

Our findings strongly suggest that hippocampal function and memory performance remain intact in recreational freedivers following repeated hypoxic exposure, ensuring that freediving, as an increasingly popular sport, does not negatively impact the hippocampus, a brain region particularly sensitive to oxygen deprivation. This raises the intriguing possibility of exploring voluntary, controlled intermittent hypoxia as a non-invasive therapeutic intervention. Whether combined with physical exercise or implemented through other protocols, moderate intermittent hypoxia holds promise as a potential strategy for enhancing neuroplasticity and cognitive resilience.