Unique Cortical and Subcortical Activation Patterns for Different Conspecific Calls in Marmosets

Abstract

The common marmoset (Callithrix jacchus) is known for its highly vocal nature, displaying a diverse range of calls. Functional imaging in marmosets has shown that the processing of conspecific calls activates a brain network that includes fronto-temporal areas. It is currently unknown whether different call types activate the same or different networks. In this study, nine adult marmosets (four females) were exposed to four common vocalizations (phee, chatter, trill, and twitter), and their brain responses were recorded using event-related functional magnetic resonance imaging at 9.4 T. We found robust activations in the auditory cortices, encompassing core, belt, and parabelt regions, and in subcortical areas like the inferior colliculus, medial geniculate nucleus, and amygdala in response to these calls. Although a common network was engaged, distinct activity patterns were evident for different vocalizations that could be distinguished by a 3D convolutional neural network, indicating unique neural processing for each vocalization. Our findings also indicate the involvement of the cerebellum and medial prefrontal cortex in distinguishing particular vocalizations from others.

Keywords: auditory; awake marmosets; conspecific vocalization; convolutional neural network; decoding analysis; fMRI.

Figure 1.

fMRI study overview. A, The sequential process of a bunched slice acquisition paradigm utilized in the study. Each acquisition cycle comprises a TA of 1.5 s followed by a TS of equal duration, collectively constituting a TR of 3 s. B, Graphical representation of the experimental task paradigm employed in the current study. Auditory stimuli with a duration of 0.6 s are randomly presented to marmoset subjects during the silent periods depicted in A with interstimulus intervals varying between 3 and 12 s, pseudorandomly chosen. C, Representation of group brain activation comparison (n = 9 marmosets) for overall auditory tasks versus the baseline. The top panels depict surface maps, providing a topographical view of cortical activations. White lines delineate regions based on the atlas from Paxinos et al. (Paxinos et al., 2012). The bottom panels show volumetric representations at different interaural (IA) levels, overlaid onto coronal slices of anatomical MR images. All surface maps are set to a threshold of z-scores below −5 and above 5, while volumetric maps are set to a threshold of z-scores below −2.3 and above 2.3, for deactivation and activation correspondingly. Cold color gradients indicate deactivation (negative values), while hot color gradients signify activation (positive values), representing the spatial distribution and intensity of neural responses during the auditory task. LH, left hemisphere; RH, right hemisphere; Aud, auditory cortex; MG, medial geniculate nucleus; IC, inferior colliculus; Amy, amygdala; RN, reticular nucleus of the thalamus; RF, the brainstem reticular formation; Cd, caudate; A1, primary auditory cortex area; ML, auditory cortex middle lateral area; AL, auditory cortex anterolateral area; CPB, caudal parabelt area; RPB, rostral parabelt area, TPO, temporoparietal–occipital area.

Figure 2.

Bilateral deactivation of the cerebellum. Volumetric representation illustrating the deactivation of the cerebellum in response to the overall auditory tasks. All volumetric maps are set to a threshold of z-scores below −2.3 and above 2.3, for deactivation and activation, respectively. Cold color gradients with negative values show deactivation, while hot color gradients with positive values denote activation, representing the spatial distribution and intensity of neural responses during the auditory task. LH, left hemisphere; RH, right hemisphere; MG, medial geniculate nucleus; Crll, cerebellum.

Figure 3.

Brain activations for the different conspecific calls versus the baseline. Group functional topologies (n = 9 marmosets) for phee (A), trill (B), twitter (C), and chatter (D) against the baseline are presented on the lateral and medial views of the right fiducial marmoset cortical surfaces. Volumetric activations at various IA levels are superimposed onto coronal slices of anatomical MR images. All activation maps are thresholded within the range of z-scores below −2.3 and above 2.3. Cool color gradients with negative values denote neural deactivation, whereas warm color gradients with positive values represent neural activation, illustrating both the spatial distribution and intensity of neural responses elicited by the auditory task. LH, left hemisphere; RH, right hemisphere; Aud, auditory cortex; MG, medial geniculate nucleus; IC, inferior colliculus; Amy, amygdala; RN, reticular nucleus of the thalamus.

Figure 4.

ROI analysis. A, The representation of neural activity patterns for each call versus the baseline on flat maps for the right hemisphere. The z-score maps are thresholded at below −2.3 and above 2.3. Warm color gradients indicate activation, while cold color gradients signify deactivation. B, Beta values analysis for the activity of each call versus the baseline across 24 ROIs. Each ROI is represented by four bars, each corresponding to the functional activity of a specific call, as displayed in the left column of the plot, wherein the bar height reflects the magnitude of activity for the corresponding ROI in response to that specific call. Significance levels of group differences in each ROI are displayed below each graph using asterisks (*p ≤ 0.05; **p < 0.01; and ***p < 0.001), displaying regions where differences reach statistical significance. In regions where differences between conditions are significant, asterisks indicating significance levels are displayed above the corresponding bars. (*p < 0.05; **p < 0.01; and ***p < 0.001). The vertical line on each bar demonstrates the SEM. CM, auditory cortex caudomedial area; A1, auditory cortex primary area; R, auditory cortex rostral area; RT, auditory cortex rostrotemporal area; CL, auditory cortex caudolateral area; ML, auditory cortex middle lateral area; AL, auditory cortex anterolateral area; RTL, auditory cortex rostrotemporal lateral area; RTM, auditory cortex rostrotemporal medial area; RM, auditory cortex rostromedial area; CPB, auditory cortex caudal parabelt area; RPB, auditory cortex rostral parabelt area; TPO, temporoparietal–occipital area; Ipro, insular proisocortex; Tpro, temporal proisocortex; STR, superior rostral temporal area; ReI, retroinsular area.

Figure 5.

Classification results from the convolutional neural network. A, After training, the CNN processes unlabeled functional maps (test data), predicting the corresponding labels. The predicted labels are then compared with the actual (true) labels to generate the confusion matrix shown here. The diagonal elements of the confusion matrix represent the correct classifications for each vocalization, while the off-diagonal elements indicate misclassifications. The chance level for this study is 25%, as there are four possible conditions. B, This figure shows the accuracy of classifying each class correctly. The accuracy for all classes is above the chance level. The vertical line of each bar demonstrates the SEM. C, Test and training accuracies during a single run of the training process are monitored. The alternating white and shaded regions represent epochs. In the top panel, the dark blue line represents the smoothed training accuracy, while the light blue line shows the actual training accuracy for each iteration. The dashed black line depicts the validation accuracy (or test accuracy) evaluated at regular intervals (frequency, 5) during training. The accuracy trends upward as the model learns from the data, with fluctuations reflecting the model's adjustments during the training. In the bottom panel, the loss (a measure of model error) during both training and validation is plotted over the same iterations. The orange line represents the smoothed training loss, the solid red line shows the actual training loss per iteration, and the dashed black line illustrates the validation loss. The overall decreasing trend in loss signifies an improvement in the model's performance, although some fluctuations occur due to the dynamic nature of training. The figure was generated using MATLAB's deep learning toolbox (R2024b).

Figure 6.

Brain activity in response to each call versus all other calls. Each row illustrates the neural response of brain regions in response to phee, trill, twitter, and chatter calls, respectively, compared with all other calls. The neural patterns are displayed on flat cortical maps for both the right and left hemispheres, along with volumetric representations at different interaural (IA) levels. All z-score maps are thresholded to display values less than −2 and >2. LH, left hemisphere; RH, right hemisphere; CPB, auditory cortex caudal parabelt area; TPO, temporoparietal–occipital area; DCN, deep cerebellar nuclei; LVIIB/LVIIIA, cerebellar lobules VIIB and VIIIA; CM, auditory cortex caudomedial area; MG, medial geniculate nucleus; SC, superior colliculus; PE, parietal area.

Figure 7.

Neural response of mPFC in response to each call versus each other calls. The volumetric representation of the neural activity of mPFC for calls on the x-axis compared with those listed on the y-axis (calls _x-axis> calls _y-axis). Results are thresholded at z-scores >2.