Hormonal contraception & face processing: Examining face gender, androgenicity & treatment duration

Previous cross-sectional studies observed differences between users and non-users of combined oral contracep- tives (COCs) in both the structure and function of the fusiform face area (FFA) related to face processing. For the present study 120 female participants performed high-resolution structural, as well as functional scans at rest, during face encoding and face recognition. Participants were either never-users of COCs (26), current first-time users of androgenic (29) or anti-androgenic COCs (23) or previous users of androgenic (21) or anti-androgenic COCs (21). Results suggest that associations between COC-use and face processing are modulated by androgenicity, but do not persist beyond the duration of COC use. The majority of findings concern the connectivity of the left FFA to the left supramarginal gyrus (SMG), which is a key region in cognitive empathy. While connectivity in anti-androgenic COC users differs from never users irrespective of the duration of COC use already at rest, connectivity in androgenic COC users decreases with longer duration of use during face recognition. Furthermore, longer duration of androgenic COC use was related to reduced identification accuracy, as well as increased connectivity of the left FFA to the right orbitofrontal cortex. Accordingly, the FFA and SMG emerge as promising ROIs for future randomized controlled trials on the effects of COC use on face processing


Introduction
Combined oral contraceptives (COCs) are the most commonly used form of hormonal contraceptives, containing an estrogen (usually ethinylestradiol) and varying levels of synthetic progestins (Sitruk-Ware, 2008). While the effects of endogenous ovarian hormones (estradiol, progesterone) on social and emotional functioning are becoming increasingly recognized (for reviews see Jones et al., 2008;Sundström Poromaa and Gingnell, 2014), studies on the effects of synthetic steroids on the brain are rare. With the increasing recognition of depression as a potential side effect of COCs (e.g. Anderl et al., 2020;Sundström-Poromaa, 2021; but see: Lundin et al., 2021), many studies focus on the effects of COCs on emotional processing (e.g. Gingnell et al., 2013;Petersen and Cahill, 2015;Engman et al., 2018;Nasseri et al., 2020). Implications for social functioning have been extensively discussed (e.g. Montoya & Bos, 2017;Porcu et al., 2019), but social processing and the underlying neural correlates were a lesser focus in COC research. In the present study, we focus on an elemental ability in social interactionsthe encoding of novel faces and the recognition of familiar faces.
Several mechanisms of action of COCs on social functioning in general, and face recognition in particular are plausible. First, COCs contain very potent synthetic steroids with a high binding affinity to estrogen and progesterone receptors (Sitruk-Ware, 2008). Thus, estrogenic and progestagenic actions are to be expected. Various studies demonstrate beneficial effects of high doses of estrogen and progesterone, as e.g., experienced during pregnancy, on social-emotional functioning and face recognition (Anderson and Rutherford, 2011;Byrne et al., 2019). Following this line of reasoning, an improvement in face recognition performance is to be expected in COC users, as was indeed observed in Pletzer et al. (2015).
Second, COCs suppress the fluctuations of endogenous ovarian hormones (Westhoff et al., 2010). Since changes in social functioning have been linked to these endogenous ovarian fluctuations (e.g. Senior et al., 2007;Jones et al., 2008), it stands to reason that the suppression of these fluctuations results in alterations of social-emotional functioning.
Third, synthetic progestins also modulate androgen receptors, as well as glucocorticoid and mineralocorticoid receptors (Sitruk-Ware, 2008). On the one hand, several progestins including the second-generation progestin levonorgestrel (LNG) and its derivates desogestrel, norgestimate and gestoden are derivates of 19-nortestosterone and therefore weak agonists of the androgen receptor. On the other hand, progestins derived from 17-acetoxyprogesterone, like chlormadinon-acetate (CMA) and cyproterone acetate (CPA), as well as progestins derived from spironolactone, like drospirenone (DSP), are anti-androgenic due to their ability to supplant androgens, like testosterone, from the androgen receptor. Since there is extensive evidence for the important role of testosterone in social-emotional functioning (see Heany et al., 2016 for a review), it seems plausible that the androgenicity of progestins modulates the effects of COCs on face recognition performance.
The androgenicity of progestins is associated with the side-effect profile of COCs (e.g. Hapgood et al., 2004) and may thus present a major confound in cross-sectional studies of COCs. Studies comparing groups of COC users to non-users regarding brain parameters and behaviour, are subject to the so-called "survivor effect" (Oinonen and Mazmanian, 2002). Women, who tolerate their COC well or show beneficial effects, are more likely to become long-term users of these contraceptives, while women who experience adverse side effects are likely to terminate COC treatment. Accordingly, women who never used hormonal contraceptives may be more suitable as a control group. Furthermore, women are more likely to use an anti-androgenic COC if they experience or expect adverse side effects with androgenic COCs, which are currently the standard of care. This confound can be reduced by focusing on first time users of either androgenic or anti-androgenic COCs. Nevertheless, simple group comparisons appear insufficient to assess the brain effects of COCs, since they are just as likely to capture brain parameters and behaviours associated with the tolerability of a certain COCs as to capture effects of the COC per se. However, while several randomized controlled trials (RTCs) of long-term COC use are underway, they are methodologically challenging and time-consuming. In the meantime, well-controlled dose-response studies examining the neural correlates of COC treatment duration may also generate hypotheses for the analyses of future RCTs. The two approaches complement each other, since cross-sectional comparisons capture changes that occur early on during the first months of treatment but stabilize over time, while associations with treatment duration arise for those effects which accumulate over time.
With regards to face processing, previous cross-sectional studies have demonstrated differences between COC users and non-users in both, structural and functional correlates, i.e. the fusiform face area (FFA) (Ganel et al., 2005;Kanwisher and Yovel, 2006). The FFA was larger (Pletzer et al., 2015) and more active during face processing (Marečková et al., 2014) in COC users compared to non-users. In both studies, the effects were more pronounced with longer COC treatment duration. These group differences in the brain were accompanied by improved face recognition performance in COC users compared to non-users. Compared to non-users, COC users were more accurate in deciding, whether they had seen a face before or not. However, several open questions remain with regard to the association of COC use to face processing.
First, it remains unclear, whether the reported associations are driven by the estrogenic, progestagenic or androgenic vs. antiandrogenic actions of synthetic steroids. While Marečková et al. (2014) did not report the type of progestin used by women in the COC group, results in Pletzer et al. (2015) point towards a positive assocition of anti-androgenic COCs, but negative asspciation of androgenic COCs to gray matter volumes of the FFA. Note however, that neither study controlled for previous contraceptive use in either the COC groups or the control group, such that the results may be confounded with the survivor effect.
Second, it remains unclear at which stage during face processing the associations to COC use are most prominent. For example, COCs could be associated to the activity of the FFA already at rest, the processing depth during the encoding of new faces, or to the accessibility of already stored face information during the recognition phase. Accordingly, in the present study we performed functional scans during all three stages, i.e. at rest, during the encoding phase and during the recognition phase.
Third, none of the previous studies has gone beyond the size and activation of well-defined, but localized regions of interest, i.e. the FFA. This area is however part of a larger brain network involved in the processing and recognition of faces (Fischer et al., 2007;Wang et al., 2016;Zhen et al., 2013). Thus, associations of COC use on the functional connectivity of the FFA during the different stages of face processing are of interest.
Forth, a consistent effect in face processing is the well-replicated own-gender bias in women (see Herlitz and Lovén, 2013 for a meta-analysis). Unlike men, women are better at recognizing faces of women than faces of men, an effect accompanied by stronger reactivity of the face-processing network to faces of women (Ino et al., 2010;Lovén et al., 2014). Accordingly, it seems relevant to assess whether COCs modulate not only overall face recognition, but also the own-gender bias in face recognition.
Finally, no study has so far addressed the reversibility of COC effects on the neural correlates of face processing. This question can be addressed by including a group of previous COC users to examine whether the results obtained in current users persist beyond the discontinuation of COC use.
To address these questions, we collected functional scans in three groups of women at rest, as well as during both, face encoding and face retrieval. The first group of women had never used COCs before and thus, did not have previous negative experiences with COCs. This group served as a control group. The second group of women were current first-time users of either an androgenic or anti-androgenic COC. Thus, they also had no previous negative experiences with other COCs and the duration of COC use was well defined for each of them. The third group were previous one-time users of either an androgenic or anti-androgenic COC. They served as a comparison group to obtain tentative evidence regarding the reversibility question. Any associations observed in current COC users, which could also be confirmed in previous COC users, suggest effects that persist after discontinuation of COC treatment.
Based on previous results from cross-sectional studies, we hypothesize an improved face recognition performance associated with enhanced activation and within-network connectivity of the face network with anti-androgenic COC use. On the contrary, impaired face recognition performance, associated with reduced activation and within-network connectivity of the face network, is expected with androgenic COC use. Exploratory analyses will be performed to address the questions during which phase (rest, encoding, recognition) and for which stimuli (men or women) associations to COC use emerge.

Participants
123 women were recruited for this study. Inclusion criteria were an age between 18 and 35 years, right-handedness, normal or corrected to normal vision, and one of the following: (i) current first-time use of a combined oral contraceptive (COC) containing ethinylestradiol (EE) in combination with levonorgestrel (LNG), drospirenone (DSP), or chlormadinon acetate (CMA), (ii) previous one-time use of a COC with known formulation, (iii) COC-naivety. Exclusion criteria were neurological, endocrine, or psychiatric disorders, regular intake of medication apart from the COC, use of multiple COCs, irregular menstrual cycles according to the criteria of Fehring et al. (2006), metal implants or brain structural abnormalities. 3 participants had to be excluded due to below-chance performance (overall accuracy < 50%) in the face recognition task.
Among the remaining 120 participants, 26 were COC-naive, 29 were current users of EE/LNG, 23 were current users of EE/DSP or EE/CMA, 21 were previous users of a COC with an androgenic progestin and 21 were previous users of a COC with an anti-androgenic progestin. Apart from Levonorgestrel (LNG), the following progestins were classified as androgenic due to being derivates of LNG or 19-nortestosterone: Desogestrel, Gestoden, Norgestimate. The following progestins were classified as anti-androgenic: Drospirenone (DSP), Chlormadinone acetate (CMA), Cyproterone acetate (CPA) and Dienogest. Demographics are displayed in Table 1. As expected, there were significant age differences between the groups (F (4111) = 7.22, p < 0.001). Current androgenic COC users were significantly younger than the other COC groups (all p < 0.03), but did not differ significantly from never users (p = 0.20). No significant age differences emerged between the other groups. Average duration of COC use was 4 years (SD = 3 years) and did not differ significantly between COC groups (F (3,93) = 1.75, p = 0.16). All except two participants, 1 who were currently or previously using COCs, had started their COC use during adolescence (≤ 21 years of age). The average age at COC start was 17 years and did not differ significantly between COC groups (F (3,93) = 1.01, p = 0.40). Furthermore, all except two previous COC users 2 were nulliparous. Previous COC users had on average discontinued their COC use 4 years ago (SD = 3 years), i.e. on average the time since discontinuation matched the duration of COC use. Time since discontinuation did not differ significantly between androgenic and anti-androgenic previous users (t (40) = − 0.91, p = 0.37). Education did not differ significantly between groups (H = 3.07, p = 0.57), but a small difference in IQ emerged due to slightly lower IQ in androgenic COC users (F (4115) = 2.60, p = 0.04). There were no group difference in relationship status (X 2 = 7.29, p = 0.12) or relationship duration (F (4,46) = 1.41, p = 0.25), i.e. COC-use was not confounded with relationship status.

Ethics statement
The study was approved by the University of Salzburg's ethics committee and conducted in accordance with the Declaration of Helsinki. All participants gave their informed written consent for study participation.

Face recognition task
Face stimuli were selected from the FACE database of the Park Aging Mind Lab (http://agingmind.utdallas.edu/facedb), as well as from the Radbound face database (Langner et al., 2010). Stimulus presentation and recording of responses were performed using software Presentation (Neurobehavioral systems, version 22.1). During the encoding phase, participants viewed 40 faces (20 men, 20 women). Each face was presented for 3 s, followed by an inter-stimulus interval (ISI) of 1.5 s. Face stimuli were interspersed by 20 null events of equal duration. During ISI's and null events, a black fixation cross was presented on a white screen. The order of stimuli and null events was randomized. To ensure participants' attention to faces, participants were asked to indicate per left or right button press on a diamond response button box (Current Designs), whether they thought the face they saw was a man or a woman. During the recognition phase, participants were presented with 80 faces (40 men, 40 women). 40 were the familiar faces they had seen during the encoding phase and 40 were novel faces. Each face was presented for 3 s, followed by an ISI of 1.5 s. Face stimuli were interspersed by 40 null events of equal duration. During ISI's and null events a black fixation cross was presented on a white screen. The order of stimuli and null events was randomized. Participants were asked to indicate per left or right button press on a diamond response button box, whether they had seen the face before or not.

MRI data acquisition
All MRI-data were acquired on a Siemens Prisma fit 3.0 Tesla scanner equipped with a 64-channel head coil. During the MRI session, a restingstate scan was acquired first, followed by the face encoding task. Between the face encoding and recognition phase, participants performed a virtual navigation and verbal fluency task, described elsewhere (Noachtar et al., 2022), as well as a high-resolution structural scan using a T1-weighted sagittal 3D MPRAGE sequence (TR = 2300 ms, TE = 2.91 ms, TI delay of 900 ms, FOV 256 mm, slice thickness = 1.00 mm, flip angle 9 • , voxel size 1.0 × 1.0 × 1.0 mm, 160 sagittal slices). In total, 30 min lapsed between the encoding and recognition phase.

Preprocessing of MRI data
Structural images were preprocessed using the voxel-based morphometry (VBM) approach implemented in the CAT12 toolbox (http://dbm.neuro.uni-jena.de/vbm/) of the SPM12 software (http ://www.fil.ion.ucl.ac.uk/spm/). Structural images were spatially registered to an anatomical template (European brains) and segmented into grey matter, white matter and CSF partitions using SPM12 tissue probability maps. Moderate strength (0.5) was chosen for local adaptive segmentation, skull stripping, final clean-up and correction of white matter hyperintensities. Bias correction was applied to control for intensity non-uniformities and segmentations were modulated by the volume change due to spatial registration [14]. Gray matter segments were smoothed using an 8 mm full width at half maximum (FWHM) Gaussian kernel.
Prior to preprocessing of functional images, voxel displacement maps (VDM) were calculated (Chang and Fitzpatrick, 1992). For resting state functionals the fieldmap was calculated from the two EPI images with opposite phase encoding using the FSL "topup" (http://fsl.fmrib.ox. ac.uk/fsl/fslwiki/TOPUP). For task based functionals the phase difference in the fieldmap was calculated using the Fieldmap Toolbox in SPM12.
Functional images were pre-processed using SPM12 standard procedures and templates. The first 6 images of each scan were discarded and the remaining scans were despiked using 3d-despiking as implemented in AFNI (afni.nimh.nih.gov). Images were realigned and unwarped using the VDM maps described above and six movement parameters were extracted. We then applied slice-timing, co-registered the functional images to the segmented structural images and normalized the functional images by applying the normalization parameters obtained during preprocessing of the structural images in CAT12. The resulting images were resampled to isotropic 3 × 3 × 3 mm voxels and smoothed with a Gaussian kernel of 6 mm.
Finally, to achieve a stricter removal of motion artifacts than possible with SPM12 realignment, the resting state functionals were subjected to the non-aggressive removal of artifactual components through the ICA-AROMA algorithm implemented in FSL (Pruim et al., 2015). The task-based functionals were additionally subjected to the Functional Image Artefact Correction Heuristic (FIACH, Tierney et al., 2016), which identifies and corrects for non-physiological noise. We filtered the images and additionally extracted six regressors of physiological noise via principal components analyses from a time-series signal-to-noise ratio (TSNR) map. After preprocessing, the following brain maps were created from the functional images: (i) Amplitude of low frequncy fluctation (ALFF) maps from resting state functionals. (ii) Activation maps from the task-based functionals. (iii) Seed-based connectivity maps from resting state and task-based functionals using the left and right FFA as seed regions.

Amplitude of low frequency fluctuation (ALFF)
ALFF maps were calculated from pre-processed resting state images using the DPABI toolbox (Yan et al., 2016). The ALFF is defined as the average square root of each frequencies power within a range of 0.01-0.08 Hz. We therefore applied a bandpass filter of 0.01-0.08 Hz to remove effects of very-low-frequency drift and high-frequency noise as caused by respiratory and heart rhythms (Zang et al., 2007). The ALFF was then obtained from the power spectrum after a Fourier transform. It is thus a measure of oscillatory activity at the resting state.

Brain activation
For task-based functionals, activation maps were obtained during subject-dependent fixed-effects first-level analysis in SPM12. For images obtained during face encoding, one regressor of interest was modelled to predict the BOLD-response to faces. For images obtained during face recognition, two regressors of interest were modelled to separately predict the BOLD response to familiar faces and novel faces. The six realignment parameters and the six physiological noise parameters obtained from the FIACH procedure were entered as regressors of no interest in the models. All regressors were obtained by convolving the duration of the event with the canonical hemodynamic response function implemented in SPM. A high pass filter cut-off was set at 128 s and autocorrelation correction was performed using an autoregressive AR(1) model (Friston et al., 2002). For each regressor, one statistical contrast was defined comparing the BOLD-response during task events to the baseline (fixation cross). This resulted in three contrast images, i.e. one for face encoding, one for the recognition of familiar faces and one for the identification of novel faces. Final activation maps were obtained by scaling the contrast images, such that the BOLD-increase during the task was estimated in relation to the oscillatory activity at rest (Kalcher et al., 2013). To that end, ALFF maps were obtained from the estimation residuals (baseline) using the DPABI toolbox as described above for resting state images. Contrast images were then divided by these ALFF maps

Functional connectivity
Seed-based connectivity analyses were performed using the CONNtooolbox (Whitfield-Gabrieli and Nieto-Castanon, 2012). Following linear detrending for white matter (WM) and cerebrospinal fluid (CSF), a band-pass filter (0.008-0.09 Hz) and motion-correction, voxel-wise connectivity maps were computed using the left and right FFA as seeds.

Statistical analysis
For statistical analyses, we first used a hypothesis driven region of interest (ROI) based approach, followed by exploratory whole brain analyses. For ROI-based analyses, the following brain parameters were extracted from the left and right FFA using FFA masks created by Pletzer et al. (2015): (i) Gray matter volumes from the gray matter segments using the get_totals script by G.Ridgeway (http://www0.cs.ucl.ac.uk/staff/ gridgway/vbm/get_totals.m). (ii) ALFF at rest (iii) Eigenvalues of BOLD-response during face encoding, recognition of familiar faces and identification of novel faces.
Behavioral and ROI-based analyses were performed with R 4.2.3. Performance and brain parameters were analysed in the context of linear mixed effects models using the lme function of the nlme package (Pinheiro et al., 2017). All models included participant number (PNr) as random factor and age and IQ as regressors of no interest. FDR-corrections were performed for each model type.
All analyses were performed on current COC users in a first step (planned analyses). If significant effects were observed in current users, the analysis was repeated in previous users to see, if the effects persisted beyond discontinuation of COC treatment (exploratory analyses).
Exploratory whole brain analyses were performed with SPM12 second level analysis. Brain maps were compared between groups using Full Factorial models and between conditions using Flexible Factorial models. In order to relate brain maps to COC treatment duration, Full Factorial models were created separately for current and past users. We entered COC treatment duration as regressor of interest and modelled its interaction with the group factor androgenicity. Regressors of no interest were age, IQ, andin case of gray matter mapstotal intracranial volume (TIV). For the second-level designs, we used an uncorrected primary threshold of p < 0.0005 and a secondary cluster-level FWEcorrected threshold of p < 0.05. The primary threshold was lowered to account for multiple comparisons in the connectivity analyses, which were repeated for two ROIs (left and right FFA). The extent threshold was set to 30 voxels. In case a cluster of significant interaction between androgenicity and pill duration emerged, eigenvalues were extracted from this cluster and partial correlations controlling for age and IQ were performed separately for androgenic and anti-androgenic COC users. In case no significant interaction between androgenicity and pill duration was observed, the main effect of pill duration was calculated across both groups. ROI-based data and scripts are openly available at https://osf. io/xc5wg/. Brain images are available from the first author upon request.

Behavioral results
Reaction times (RT) for correct responses did not differ significantly between faces of men and women, irrespective of whether they were familiar or novel (both F (1225) < 1.52, both p > 0.21, both η 2 < 0.02).
There was no significant interaction between current COC use and face gender for either familiar or novel faces (both F (2225) < 0.14, both p > 0.87, both η 2 < 0.01). RT did not differ significantly between current COC users and never-users irrespective of face novelty (both F < 2.25, both p > 0.11, both η 2 < 0.06).

Neuroimaging results
Both face encoding and face recognition elicited significant activation in the bilateral occipital and fusiform gyri, pre-and post-central gyri, supplementary motor area (SMA), Insulae and inferior frontal gyri. Face encoding additionally elicited significant activation in the bilateral hippocampi ( Fig. 2A in red). Face recognition additionally elicited significant activation in the bilateral superior parietal lobes (SPL) and bilateral dorsolateral prefrontal cortices (DLPFC) (Fig. 2B, familiar faces in red, novel faces in yellow).
During face encoding and face recognition alike, faces of women elicited significantly stronger BOLD-response than faces of men in the bilateral occipital and fusiform gyri. During face encoding, stronger activation for faces of women extended to the left pre-and postcentral gyri and right DLPFC (Fig. 2C in red). During recognition of familiar faces, stronger activation for faces of women extended to the superior parietal lobes and bilateral DLPFC (Fig. 2D in red). During face encoding, faces of men elicited significantly less deactivation in default mode areas and significantly stronger activation in the right pre-and postcentral gyri compared to faces of women ( Fig. 2C in green). Likewise, during face recognition (both familiar and novel faces) faces of men elicited less deactivation in the anterior cingulate cortex (ACC) than faces of women ( Fig. 2D in dark green). In addition, novel faces of men elicited stronger BOLD-response in the SMA and bilateral Insulae than novel faces of women (Fig. 2D in dark green).
During face encoding and face recognition alike, the FFA showed significant positive connectivity to all task-positive areas ( Fig. 2A-B) and significant negative connectivity to default mode areas. Only few connectivity differences between faces of men and women emerged (results not shown). During encoding, connectivity to the left FFA was significantly stronger for faces of women compared to faces of men in the right postcentral gyrus ([15, − 31, 76] [57, − 43, 28], 58 voxels, T = 4.43, p FWE < 0.001). Additionally, connectivity to the left FFA was stronger for faces of men compared to faces of women in the right anterior Insula ([33, 17, 4], 36 voxels, T = 4.77, p FWE = 0.005). No connectivity differences emerged between novel faces of men and women.

COCs and brain structure
Gray matter volumes (GM) of the FFA did not differ significantly between current COC users and never-users, irrespective of hemisphere (both F (2,75) < 0.97, both p > 0.38, both η 2 < 0.10). However, there was a significant interaction between androgenicity and the duration of current COC use for gray matter volumes of the FFA (F (1,45) = 6.05, p FDR

COCs and the face processing network at rest
Irrespective of hemisphere and androgenicity, the amplitude of low frequency fluctuations (ALFF) in the FFA did not differ significantly between never-users and COC users and was not associated to the duration of COC use (all F < 1.62, all p > 0.20, all η 2 < 0.02).
Exploratory whole brain analyses revealed that at rest, both the left and right FFA showed significant positive connectivity to all taskpositive areas (compare Fig. 4A-B) and significant negative connectivity to default mode areas. Anti-androgenic COC users showed significantly stronger connectivity between the left FFA and the left SMG than never-users ([− 54, − 52, 46], 83 voxels, T = 5.46, p FWE < 0.001; Fig. 4 left, yellow). Previous anti-androgenic COC users had values between Fig. 2. Brain activation during face encoding (A-B) and recognition (C-D). A. Overall brain activation patterns during face encoding (red) and fusiform face area (FFA, blue circles). B. Overall brain activation pattern during the recognition of familiar faces (yellow) and identification of novel faces (red), as well as fusiform face area (FFA, blue circles). C. Comparison of activation patterns during encoding of faces of men and women. Stronger activation for faces of women compared to faces of men is depicted in red. Stronger activation for faces of men compared to faces of women is depicted in green. D. Comparison of activation patterns to familiar and novel faces of men and women during recognition. Stronger activation for faces of women compared to faces of men is depicted in red for familiar faces and yellow for novel faces. Stronger activation for faces of men compared to faces of women is depicted in dark green for familiar faces and light green for novel faces. never-users and current users, but differences to never-users were nonsignificant (b = 0.42, SE b = 0.28, t (113) = 1.49, p = 0.14). Neither in this area, nor in any other area at the whole brain level, was connectivity of the left or right FFA related to the duration of COC use.

COCs and the face network during face encoding
Irrespective of face gender and hemisphere, we observed no significant differences between COC users and never-users and no association between the duration of current COC use and BOLD-response in the FFA (all F < 3.96, all p FDR > 0.06, all η 2 < 0.07).
Exploratory whole brain analyses revealed that -as in the resting state -connectivity of the left FFA was significantly stronger in antiandrogenic COC users compared to never users to the left SMG ([− 51,− 49, 49], 64 voxels, T = 5.13, p FWE < 0.001; compare Fig. 4 right, red) irrespective of face gender. During encoding, this effect was also confirmed in previous users (b = 0.91, SE b = 0.28, t = 3.27, p = 0.001).
Smaller clusters of stronger connectivity to the left FFA in antiandrogenic COC users compared to never-users emerged in the SMA At the whole brain level, neither activation nor connectivity showed associations to the duration of current COC use.

COCs and the face network during face recognition
Neither during the recognition of familiar faces, nor during the identification of novel faces did we find significant differences between Fig. 4. Connectivity of the left FFA to the left SMG at rest (left, yellow) and during face encoding (right, red). While never-users showed negative connectivity between the left FFA and the left SMG (red), these areas were disconnected in anti-androgenic COC users.

Fig. 5. Association between duration of COC-use and connectivity between the left FFA and bilateral SMG during recognition of familiar (A, B: orange) and identification of novel faces (C-D, B: red).
Longer duration of androgenic COC use (blue), but not anti-androgenic COC-use (red) was associated with less connectivity between the left FFA and the SMG. never-users and COC users in the activation of the FFA irrespective of hemisphere (all F < 1.27, all p FDR > 0.41, all η 2 < 0.03) or any other area at the whole brain level and no group differences in the connectivity of the FFA.
Neither during the recognition of familiar faces, nor during the identification of novel faces was the BOLD-response in the FFA associated with the duration of COC use irrespective of androgenicity and hemisphere (all F < 4.77, all p FDR > 0.09, all η 2 < 0.04).
Furthermore, during the recognition of familiar faces, both the left and right FFA showed stronger connectivity to the right orbitofrontal cortex (OFC) with longer duration of COC use irrespective of face gender (left: 56,30 voxels,T = 4.46,p FWE = 0.014;right: [36,41,, 82 voxels, T = 5.50, p FWE < 0.001; Fig. 6). Although the interaction with androgenicity was non-significant at the whole-brain level, extraction of eigenvalues revealed that this association was strongly driven by androgenic COC users (androgenic: both r > 0.73, both p < 0.001; anti-androgenic: both r < 0.34, both p > 0.13; compare Fig. 6). These associations were also not confirmed in previous COC users.

Discussion
The aim of the current study was to evaluate associations between COC use and face processing in a two-fold approach. First, we assessed differences between first-time users of COCs and never-users to capture effects that arise during the first months of treatment, but stabilize early on. Second, we assessed associations between face processing and the duration of COC use to capture effects that accumulated over time in a dose-response fashion. In all analyses we differentiated between androgenic and anti-androgenic COCs and explored (i) during which phase of face processing the effects occurred, (ii) whether the effects were modulated by face gender and (iii) whether the associations persisted beyond the duration of COC treatment in previous users of COCs.
Associations to COC use emerged mostly with respect to the connectivity between the left FFA and the left SMG and irrespective of face gender. While in anti-androgenic COC users this connection showed reduced anti-correlation as compared to never-users irrespective of treatment duration, in androgenic COC users this connection showed increased anti-correlation with longer duration of COC use. Furthermore, association to the duration of androgenic COC use also emerged with respect to performance, brain structure and connectivity to the orbitofrontal cortex.
Before discussing these results in more detail, we would like to put overall activation and connectivity patterns into perspective. Overall activation patterns during face encoding and face recognition showed a strong overlap and were in line with previously observed face positive networks in women (Fischer et al., 2007;Mather et al., 2010). Connectivity patterns of the FFA during face encoding and face recognition strongly overlap with the face positive network, supporting the notion that the face network is a rather encapsulated network with strong within-network connectivity, but weak connections to other networks Zhen et al., 2013).
As repeatedly demonstrated for women, we observed better accuracy for faces of women than men in the absence of differences in response time (see Herlitz and Lovén, 2013 for a meta-analysis). This difference in performance was accompanied by stronger activation of task-positive areas, including the FFA, for faces of women compared to faces of men during all phases of face processing (compare Ino et al., 2010;Lovén et al., 2014, but see Fischer et al., 2004. Indeed, accuracy during face recognition was positively related to activation of the bilateral FFA during face recognition (all r > 0.29, all p < 0.001), though no associations to the activation of the bilateral FFA during face encoding were observed (both r < 0.12, both p > 0.07). Interestingly, and not investigated in previous studies, connectivity differences between faces of men and women emerged mostly during the encoding phase and were hardly observed during the recognition phase. Specifically, during the encoding phase, we observed stronger connectivity between the right FFA and OFA for faces of women compared to faces of men. A strong connection between the OFA and FFA has previously been suggested as a pre-requisite for successful face processing since it is lacking in participants with impaired face recognition (Steeves et al., 2006;Lohse et al., 2016;Maher et al., 2019;Cohen et al., 2019). During the recognition of familiar faces, the most prominent connectivity difference between faces of men and faces of women was observed in the right SMG. The right SMG has been repeatedly described as an empathy area with the primary function to decouple the perception of oneself from the perception of others (Lawrence et al., 2006;Silani et al., 2013;Hoffmann et al., 2016;Fig. 6. Association between duration of COC-use and connectivity between the bilateral FFA and the right orbitofrontal cortex. Longer duration of androgenic COC use (blue), but not anti-androgenic COC-use (red) was associated with stronger connectivity between the bilateral FFA and the right orbitofrontal cortex (OF). Wada et al., 2021). It can be speculated that the own-gender bias is the result of stronger identification of women with faces of other women compared to faces of men, which would be consistent with the reduced activation of the right SMG for faces of women. Despite these differences in overall connectivity patterns during the processing of faces of women and faces of men, face gender did not modulate any of the associations between COC use and face processing.
Turning to the modulation of face processing by COC use, three general aspects are noteworthy. First, all but one of the results observed in current COC users, were non-significant in previous COC users, suggesting that the associations to COC use reversed after discontinuation of COC use. Second, all results were modulated by the androgenicity of the progestin contained in the COC, pointing to an important role of androgen receptor modulation for face processing. Third, for antiandrogenic COC users we observed group differences, but no associations to the duration of COC use, while the opposite pattern was observed in androgenic COC users. This suggests that for antiandrogenic COCs the differences emerged during the first months of COC use and stabilized with time, while for androgenic COCs the differences accumulated over time.
Irrespective of these differences in temporal dynamics, the same core network characteristic emerged as related to anti-androgenic and androgenic COC use, i.e. the connectivity between the left FFA and the left SMG. While associations of anti-androgenic COC use to this connection already emerged at rest and persisted into the encoding phase, associations to androgenic COC use were observed during the recognition phase. It can be speculated that these phase differences between androgenic and anti-androgenic COC users relate to the different temporal dynamics between the groups. While the more stabilized changes in anti-androgenic COC users appear to be taskinvariant, the accumulative changes in androgenic COC users appear more flexibly during specific tasks.
In never-users, the BOLD-response fluctuations of the left FFA were anti-correlated with the BOLD-response fluctuations of the left SMG, which was part of the task-negative network. This anti-correlation was reduced in anti-androgenic COC users irrespective of face gender. While anti-androgenic COC use results in less negative connectivity between those areas, androgenic COC-use appears to gradually decrease connectivity between those areas, resulting in more negative connectivity for longer androgenic COC use. In line with this observation, altered connectivity of the supramarginal gyrus has also been reported upon testosterone administration (Votinov et al., 2020). While the right SMG has been ascribed a role in distinguishing self and others in individual studies as described above, a meta-analysis on empathy tasks suggests an important role of the left SMG in cognitive empathy (Kogler et al., 2020). In addition, the left SMG plays an important role in episodic memory and habituation, though this has mostly been explored for verbal information (e.g. Vines et al., 2006;Deschamps et al., 2014;Guidali et al., 2019). In any case, the SMG is part of the fronto-parietal network, which is involved in the selection of sensory information by attention (Ptak, 2012). It is thus plausible that the stronger recruitment of this area by the left FFA supports the recognition of faces.
In previous studies we had observed improved face recognition accuracy in anti-androgenic COC users compared to non-users, which were mostly previous users of androgenic COCs (Pletzer et al., 2015). Such performance differences were not observed in the current study, but the duration of androgenic COC use was related to decreased accuracy in the identification of novel faces, supporting the notion, that increased anti-correlation between the left FFA and the left SMG results in performance decreases. The fact that performance was reduced with longer androgenic COC use is in line with previous results on sex differences in face recognition (Herlitz and Lovén, 2013). In the light of this interpretation also the larger FFA volumes with longer androgenic COC use appear plausible, given that mendespite lower face recognition performancehave larger FFA volumes compared to women (e.g. Pletzer et al., 2010).
Finally, longer androgenic COC use is related to increased connectivity between the bilateral FFA and the right orbitofrontal cortex. The orbitofrontal cortex has been repeatedly discussed as part of social, emotional and reward-related brain networks (Rolls et al., 2004;Stalnaker et al., 2015), and has been assumed to be responsible for assigning value to social-emotional stimuli (Padoa-Schioppa, 2009). Furthermore, multiple studies have demonstrated a testosterone dependent modulation of orbitofrontal functioning, including both activation and connectivity patterns during social processing, specifically during threat assessment (Mehta and Beer, 2010;van Wingen et al., 2010;Spielberg et al., 2015). The combination of a reduced FFA connectivity to the SMG, but increased FFA connectivity to the orbitofrontal cortex with longer androgenic COC use suggests that over time androgen receptor modulation may slightly shift women's responses to social cues. However, such functional inferences from brain connectivity patterns are highly speculative.
As a limitation it should be noted that the moderate group sizes and number of different outcome measures (6 in total) results in small power for the detection of small and moderate effects. Accordingly, only the largest and most robust effects were reported in this paper. Also, the majority of findings stem from exploratory whole-brain analyses, while region of interest based approaches have yielded mostly non-significant findings.
In summary, COCs relate to face processing at multiple levels, including performance, brain structure, (right) FFA activation and connectivity of the (left) FFA to the SMG during all stages of face processing. Given the correlational nature of this study, no causal inferences can be made, but the FFA and SMG as well as their respective interconnections emerge as promising ROIs for future randomized controlled trials on the effects of COC use on face processing. The present study suggests, that associations between COC use and face processing are modulated by the androgenicity of the COC and do not persist beyond the duration of COC use.

Declaration of Competing Interest
The authors declare no conflicts of interest.