AI processing of fMRI data knows with a 94+ pct accuracy when are the subjects seeing sexual images

The Brain Activation-Based Sexual Image Classifier (BASIC): A Sensitive and Specific fMRI Activity Pattern for Sexual Image Processing. Sophie R van ’t Hof, Lukas Van Oudenhove, Erick Janssen, Sanja Klein, Marianne C Reddan, Philip A Kragel, Rudolf Stark, Tor D Wager. Cerebral Cortex, bhab397, December 16 2021. https://doi.org/10.1093/cercor/bhab397

Abstract: Previous studies suggest there is a complex relationship between sexual and general affective stimulus processing, which varies across individuals and situations. We examined whether sexual and general affective processing can be distinguished at the brain level. In addition, we explored to what degree possible distinctions are generalizable across individuals and different types of sexual stimuli, and whether they are limited to the engagement of lower-level processes, such as the detection of visual features. Data on sexual images, nonsexual positive and negative images, and neutral images from Wehrum et al. (2013) (N = 100) were reanalyzed using multivariate support vector machine models to create the brain activation-based sexual image classifier (BASIC) model. This model was tested for sensitivity, specificity, and generalizability in cross-validation (N = 100) and an independent test cohort (N = 18; Kragel et al. 2019). The BASIC model showed highly accurate performance (94–100%) in classifying sexual versus neutral or nonsexual affective images in both datasets with forced choice tests. Virtual lesions and tests of individual large-scale networks (e.g., visual or attention networks) show that individual networks are neither necessary nor sufficient to classify sexual versus nonsexual stimulus processing. Thus, responses to sexual images are distributed across brain systems.

Keywords: erotic images, machine learning prediction model, multivariate analysis, neuroimaging, sexual stimuli processing, support vector machine classification

Discussion

Sexual stimulus processing is a core component of human affective and motivational systems, and part of a fundamental repertoire of motivations conserved across nearly all animal species. Previous work using sexual stimuli has made important advances (e.g., Georgiadis et al. 2006; Walter et al. 2008b; Abler et al. 2013; Borg et al. 2014; Stark et al. 2019), but these studies have generally included small sample sizes and have focused on characterizing responses in individual brain regions using standard brain-mapping approaches. Findings have been variable across studies (for meta-analyses, see Stoléru et al. 2012; Poeppl et al. 2016), and it remained unclear whether brain responses to sexual stimuli are robustly and reproducibly different from responses to nonsexual positive or negative affective stimuli.

Here, we employed a multivariate predictive model grounded in population-coding concepts in neuroscience (Pouget et al. 2000; Shadlen and Kiani 2007; Kragel et al. 2018) and systems-level characterization, based on growing evidence that various psychological processes are grounded in distributed networks rather than local regions or isolated circuits (Kamitani and Tong 2005; Kuhl et al. 2012; Arbabshirani et al. 2017). We identified a generalizable pattern of brain responses to sexual stimuli whose organization is conserved across individual participants, but which is distinct from responses to other conceptually related (nonsexual) affective images. We used cross-validated machine learning analyses to identify a brain model, which we termed the BASIC model (for purposes of sharing and reuse), that can classify sexual from neutral, positive, and negative affective images with nearly perfect accuracy in forced-choice tests, including an independent validation cohort tested on a different population (US vs. Europe), scanner, and stimulus set from those used to develop the model. Together with previous smaller-sample analyses that differentiate multivariate brain responses to romantic or sexual stimuli from responses to other types of affective and emotional events (Kassam et al. 2013; Kragel et al. 2019), our results suggest that sexual stimuli are represented by a relatively unique brain “signature” that is not shared by other types of affective stimuli.

Furthermore, our virtual lesion analysis suggests that the classifications of sexual versus neutral/affective conditions are not solely due to differences in visual or attention processing, as predictions are intact even leaving out large-scale cortical networks devoted to attention and vision. In addition, the spatial scale evaluation demonstrates that whole-brain level classification (both voxel- and parcel-wise) shows the highest model performance compared with individual large-scale network parcels. The BASIC model shows effects not only in subcortical but also in cortical areas, in line with previous human (for meta-analyses, see Stoléru et al. 2012; Poeppl et al. 2016) and animal research (for meta-analysis, see Pfaus 2009). From a basic biological perspective, this might be surprising. Evolutionarily relevant key features of sexual signals in nonhuman primates may include sex calls, pheromones, and the presentation of genitals. The sexual signals presented here are, in comparison, highly complex visual scenes containing a variety of sexual content, triggering valuation processes accompanied by neural activity on the cortical level. Even though our and previous research shows strong evidence for large cortical involvement, there still seems to be a bias in picking brain areas for region of interest (ROI) analyses toward subcortical regions. This is reflected in the neurosynth “sexual” brain map, based on an automated meta-analysis that includes coordinates from a priori ROI analyses. For example, the study with the highest loading on the term “sexual” in neurosynth (Strahler et al. 2018) used ROI analyses that included almost exclusively subcortical areas.

Many types of validation are beyond the scope of this study, but we were able to provide validation of several key elements. First is the application to a new cohort with different population characteristics, equipment, and paradigm details, with large effect sizes for sexual versus nonsexual affective images. Second, we investigated the effects of globally distributed signal in white matter and ventricle spaces, which can capture complex effects of head movement and task-correlated physiological noise and have been found to drive some multivariate predictive models in the past. Lack of relationships with these non–gray matter areas, along with significant contributions to the model in known affective/motivational systems, increases confidence that the model is driven by neuroscientific relevant systems. Third, we investigated whether the model showed differential effects for male versus female subgroups or varied with age. It did not, supporting the notion that despite individual differences there is a generalizable brain response across individuals (note, this study did not include nonheterosexual, noncis individuals, and individuals of different age groups). This is in line with the findings of previous neuroimaging meta-analyses that revealed common “unisex” brain responses to sexual stimuli (Poeppl et al. 2016; Mitricheva et al. 2019).

Interpretation of a machine learning–based model is complex because the classification is not explained by one region or network, but by a combination across regions. One set of regions may encode one aspect, for example the positive valance aspect, another set may encode the arousal aspect, and yet another set the concept of personal closeness. All these sets then jointly contribute to the overall discrimination of sexual from general affective images. The studies we analyzed do not have sufficient information to link specific brain areas to specific component processes underlying response to sexual images, but we do evaluate our model in light of previous neuroscientific literature here to examine the neurobiological plausibility of the model (Kohoutová et al. 2020).

Brain areas included in the BASIC model are also present in the most recent meta-analytical model of brain responses to sexual stimuli (Stoléru et al. 2012), although the BASIC model presents a more comprehensive and precisely specified set of hypotheses about which voxels, with which relative activity pattern across them, to test and validate in future studies. In terms of resting-state networks (see Fig. 5), we see positive and negative weight effects emerge: negative weights (relative decreases in activity associated with sexual image processing) in somatomotor networks and positive weights (relative increases) in dorsal and ventral attention networks. In addition, weights in the default mode network (DMN) are near-zero when averaging across the entire DMN. However, when looking at default mode subnetworks, DMN A (ventral medial PFC and posterior cingulate areas) shows strong positive weights, whereas DMN C (hippocampal and more posterior occipital areas) shows strong negative weights. This relates to previous research linking DMN to drug, gambling, and food craving and their regulation, which generally involve DMN A regions, and the vmPFC and NAc in particular (Hare et al. 2009; Kober et al. 2010; Hutcherson et al. 2012; Kearney-Ramos et al. 2018; Aronson Fischell et al. 2020; Schmidt et al. 2020). Both these areas are involved in the BASIC model, in line with previous research linking these areas to sexual stimuli. For instance, previous studies have reported significant vmPFC activation during sexual compared with monetary rewards (Schmidt et al. 2020), and neural reactivity to sexual stimuli in the NAc was positively correlated with sexual arousal ratings (Klein et al. 2020). In addition, activation of both regions to food and sex cues has been found to predict subsequent risky sexual behavior (Demos et al. 2012).

The BASIC model included positive weights (a higher likelihood that the image was sexual with increasing activity) in several additional regions thought to be important for sexual responses: the hypothalamus, amygdala, somatomotor cortices, and insula. The hypothalamus is a small area near the ventricles and sinus spaces, and this likely introduces substantial variability. A preliminary study by Walter et al. (2008b) using ultra high-resolution imaging at 7 T, which is likely to have superior ability to detect hypothalamic activity, found signal dropouts in ventral subcortical structures such as the hypothalamus. Similar dropout may have limited sensitivity in the hypothalamus in this study. However, findings of hypothalamic activation in sexual stimulus processing have varied across studies. The meta-analysis by Stoléru et al. (2012) found that 37.8% of studies reported hypothalamic responses to visual sexual stimuli. A motivational role for hypothalamus is included in the model of Stoléru et al. (2012) as well, although this has only been found in animal studies. Responses of the hypothalamus have been related to the regulation of autonomic responses, and in particular the physiological aspect of sexual arousal (Ferretti et al. 2005). Here, we did not measure genital responses and can therefore not know if the stimuli triggered a physiological response. Further research, including genital response measures, could therefore shed more light on the role of the hypothalamus in sexual behavior.

The amygdala and somatomotor cortices are part of the emotional component of the model of sexual stimuli processing by Stoléru et al. (2012). All these show positive weights in the BASIC model. Within the amygdala, positive weights were found in the corticomedial division, in contrast from emotions more generally, which most often show central nucleus and sometimes basolateral activation (Wager et al. 2008; Yarkoni et al. 2010). The insula has previously been reported to sex, food, and drug craving (Pelchat et al. 2004; Yokum et al. 2011; Murdaugh et al. 2012; Tang et al. 2012), as well as interoception (Paulus and Stewart 2014). The insula is, however, large and heterogeneous region. Within the insula, the BASIC model included positive weights in two areas: the right ventral anterior insula and posterior insula PoI2 (from Glasser et al. 2016). The posterior insula is held to be important for somatosensory representations, multisensory information, and pleasant touch (Olausson et al. 2002; Cera et al. 2020). The anterior insula seems to play a role in visceral information processing and subjective feelings (Craig 2002; Uddin 2015). However, the ventral anterior insula is distinct from the dorsal anterior insula identified in most studies and has stronger associations with ventromedial prefrontal and subcortical structures including the amygdala, and functional associations with emotion and gustation (Chang et al. 2013; Wager and Barrett 2017).

In addition, another subcortical region little discussed in the Stoléru et al. (2012) meta-analysis but involved in the BASIC model is the midbrain periaqueductal gray (PAG). The PAG is best-known for its role in pain and defensive behaviors, but animal literature also shows effects of lesions on sexual behavior (Lonstein and Stern 1998), particularly lordosis. Many PAG neurons express estrogen receptors, and areas where these neurons are concentrated are targeted by inputs from the hypothalamus (Bandler and Shipley 1994). This and related pathways through the PAG are thought to be involved in sexual readiness (Holstege and Georgiadis 2004). In humans, nearby areas of the midbrain are activated during male ejaculation (Holstege et al. 2003; Georgiadis et al. 2009), though precise localization is difficult, and other studies have related human PAG activity more to bonding than sex (Ortigue et al. 2010).

The overlap of areas in BASIC model with some drug- and food-cue reactivity studies, but not others, suggests that different types of appetitive stimuli and responses may activate dissociable systems in some cases. Exploring these differences in depth is beyond the scope of this study but very interesting for future studies. An interesting next step, for instance, would for example be to test BASIC model on a different set of rewarding stimuli.

Thus, future validation for BASIC model can involve testing it on many other types of stimuli but our work already shows that based on brain data, we can distinguish sexual from general affective processing. Previous research has associated sexual stimuli with positive affect, as the wide use of IAPS, where sexual images are placed under positive affect, indicates. This strong link between positive affect and sexual stimuli might be the result of the assessment method of affect. Most studies have used a bipolar scale (i.e., negative to positive affect/valence) but when using to separate unipolar scales, both positive and negative affect have been reported during sexual stimuli (Peterson and Janssen 2007). Here, we show that the BASIC model has the highest cosine similarity with the sexual condition as expected (see Supplementary Figure 1) but shows more cosine similarity to the negative condition in Study 1 than the positive condition. Hence, in line with previous research, this work therefore demonstrates that the intuitive link between positive affect and sexual stimuli is much more complicated.

In addition to the link between sexual stimuli and general affect, we were able to gain some insight into whether the BASIC model captures general arousal or valence. To examine the role of general arousal and valence in the prediction of BASIC model, we adopted two strategies. First, we measured valence and arousal in Study 1 and performed a sensitivity analysis, testing whether the BASIC model was sensitive to valence and arousal of nonsexual images. Second, we applied the BASIC model to an independent test dataset (Study 2), in which sexual and nonsexual images were matched on valence and arousal. Regarding the first strategy, the self-reported data showed that sexual images had a significantly lower arousal than the negative images, but significantly higher BASIC responses. In addition, positive images had a higher valence than neutral or negative images but did not produce higher BASIC responses. Regarding the second strategy, the BASIC model responded more strongly to sexual than nonsexual images matched on valence and arousal and did not respond to either positively or negatively valenced nonsexual images. In addition, the strong classification performance was replicated in both Study 1 and Study 2 despite differences in the content of sexual images (Study 1 showed explicit sex scenes with couples, whereas Study 2 showed both clothed and naked couples and individuals) and likely general arousal levels. Together, these findings indicate that it is not sensitive to general arousal and valence per se but is instead sensitive to sexual content. This is in line with a previous study by Walter et al. (2008a), demonstrating that during a sexual stimulus, activation patterns modulated by general emotional arousal differed from activation patterns modulated by sexual stimulus intensity.

Sexual stimuli have often been used by researchers to study sexual arousal, although it is unclear if a state of sexual arousal is elicited by short visual sexual stimuli and therefore whether it was present during conditions used in BASIC model. In Study 1, the sexual images consisted of heterosexual couples engaged in sexual interactions, and self-reported sexual arousal was significantly higher in the sexual conditions versus the other conditions. Based on these results, we might suspect that the images, although presented for a short duration, might have induced a certain level of sexual arousal, at least at the subjective self-report level. However, in Study 2, participants were presented not only with couples, but also sexual or romantic images of an individual man or woman. Assuming that not all participants were bisexual, participants were presented with sexual images depicting both individuals consistent and inconsistent with their preferred sex. Thus, even though the images might not have induced high levels of sexual arousal in all participants, we can distinguish sexual from general affective processing in the brain, which was the aim of our study. For future research, it would be interesting to examine whether the BASIC model can also differentiate between longer visual sexual stimuli, whether sexual arousal is more likely to be induced by long than short-duration, and whether the BASIC model responds to sexual stimuli of other modalities, for example, sensory (genital stimulation), cognitive (fantasy), or auditory.

Ponseti et al. (2012) conducted one of the few studies on sexual stimulus processing that used multivariate analysis. They classified preferred and nonpreferred (e.g., child nudity vs. adult nudity) sexual stimuli based on brain data in participants with and without pedophilia using nude frontal images of adults and children. This classification might be more linked to sexual arousal, although it is still hard to evaluate whether sexual arousal was induced. In order to gain additional insight into sexual arousal specifically, future research could examine if sexual arousal is elicited during sexual image presentation, and to identify the brain processes generating it, multivariate analysis could be used to predict sexual arousal ratings based on brain data collected during sexual image presentation. In our study, this was not possible due to a lack of within-subject variability in the sexual arousal ratings during the sexual image blocks. Parada et al. (2016) presents large variability of sexual arousal ratings and, using a parametric modulation analysis, found various subregions of the parietal cortex that showed significant changes in activation corresponding to the degree of self-reported sexual arousal with no gender differences. Future studies could further examine role of the parietal cortex in the subjective experience of sexual arousal.

Besides self-reported sexual arousal, genital responses are often assessed in psychophysiological studies to examine sexual arousal (Rosen and Beck 1988; Janssen and Prause 2016). The assessment of genital response in neuroimaging studies is sparse (Arnow et al. 2009; Parada et al. 2018). Parada et al. (2018) presents several brain regions (supramarginal gyri, frontal pole, lateral occipital cortex, and middle frontal gyri in men; same regions plus the ACC/PCC, right cerebellum, insula, frontal operculum, and paracingulate gyrus) to be correlated with changes in genital response, with a stronger brain–genital relation in women compared with men in several regions. Assessment of genital response during fMRI research could improve our understanding of the interaction between brain and genital and the gender differences between this interaction. In addition, multivariate analysis could be used to predict genital arousal levels and self-reported sexual arousal based on brain data, and these patterns could be compared. This type of study design would allow for a mediation analysis, which could give more insight into the brain organization by examining the distributed, network-level patterns that mediate the stimulus intensity effects on sexual arousal (Geuter et al. 2020).

A limitation of this study is that although the BASIC model can accurately classify sexual and nonsexual images with forced choice tests, we did not identify one absolute threshold that could be used as a quantitative measure across studies. Future studies thus have to establish a threshold in a study-specific manner and make relative comparisons across conditions within-study, which is a limitation. However, we do show that the BASIC model can be generalized to individuals studied in other research centers with forced choice tests, though the absolute scale of the response is likely to vary across studies as a function of scanner field strength, signal-to-noise ratio, and other signal properties.

To summarize, in this study, we applied multivariate neuroimaging analyses to investigate sexual stimulus processing in the brain. This approach allowed for the development of the BASIC model, which can accurately classify sexual versus neutral and positive and negative affective images in two separate datasets, consisting of different types of sexual stimuli and individuals. The BASIC model includes a precisely specified pattern of cortical and subcortical areas, some of which have received relatively little attention in the literature on human sexual responses (e.g., cortical networks). Some may be shared across other appetitive responses (e.g., vmPFC and NAc for drug cues), but the BASIC model may also diverge from studies of other appetitive responses as well (e.g., in the insula). The work gives insight into the complex processing of sexual stimuli and supports the notion that processing sexual stimuli is a neurologically complex, potentially unique mental event that involves multiple networks distributed in the brain. There are many avenues open for future validation and further development, such as testing the BASIC model to nonsexual rewarding stimuli or sexual stimuli of other modalities, and linking the work to sexual arousal.