Functional parcellation of mouse visual cortex using statistical techniques reveals response-dependent clustering of cortical processing areas

Abstract

The visual cortex of the mouse brain can be divided into ten or more areas that each contain complete or partial retinotopic maps of the contralateral visual field. It is generally assumed that these areas represent discrete processing regions. In contrast to the conventional input-output characterizations of neuronal responses to standard visual stimuli, here we asked whether six of the core visual areas have responses that are functionally distinct from each other for a given visual stimulus set, by applying machine learning techniques to distinguish the areas based on their activity patterns. Visual areas defined by retinotopic mapping were examined using supervised classifiers applied to responses elicited by a range of stimuli. Using two distinct datasets obtained using wide-field and two-photon imaging, we show that the area labels predicted by the classifiers were highly consistent with the labels obtained using retinotopy. Furthermore, the classifiers were able to model the boundaries of visual areas using resting state cortical responses obtained without any overt stimulus, in both datasets. With the wide-field dataset, clustering neuronal responses using a constrained semi-supervised classifier showed graceful degradation of accuracy. The results suggest that responses from visual cortical areas can be classified effectively using data-driven models. These responses likely reflect unique circuits within each area that give rise to activity with stronger intra-areal than inter-areal correlations, and their responses to controlled visual stimuli across trials drive higher areal classification accuracy than resting state responses.

Fig 1. Experimental setup.

A) Diagrammatic representation of a mouse prepared for wide-field imaging. B) Diagrammatic representation of the custom-made one-photon wide-field imaging setup along with the display screen. C) Display configuration relative to imaging the left cortex. D) Visual cortex map showing the sign of visual field representations along with areal boundaries for six visual areas used in this study.

Fig 2. Pipeline for supervised classification of visual areas.

A) Block diagram for supervised classification of visual cortex. B) The pixels chosen for training the classifiers are shown as black dots. C) Result of classifying visual cortex using the supervised GMM classifier. Boundaries in black denote the ground truth.

Fig 3. Analysis of visual cortex responses to different stimuli using supervised GMM classifiers.

A, B) The boundaries obtained by classifying all the pixels using the GMM classifier. Each color represents the visual area identified by the classifier and the black boundary within the cortex corresponds to ground truth retinotopic boundaries. The values within the bracket denote the classification accuracy. In A, the results are compared across different visual stimuli. The title of the plot indicates the visual stimuli shown to the mice. In B, the supervised classifier is verified to be consistent across different mice for natural movie stimuli. C) Results on resting state responses for two mice. D) Pixels selected for training the supervised model are limited to center x% of the radius of the visual area. This x% is mentioned as sample radius in the title of the plots in the first row of D, and the pixel used for training the supervised model is shown as black dots. The corresponding classification boundaries are shown in the second row of D, and the “ACC” values denote the accuracy.

Fig 4. Confusion matrices for test data obtained using supervised classifier.

The diagonal values denote the precision (in %) of each class. Off-diagonal values denotes the false prediction rate (in %) for the predicted class given the actual class. A) Confusion matrix obtained using responses of Mouse M1 and Natural Movie stimuli. B) Confusion matrix obtained using the Cre-line Emx1-IRES and Natural Movie 3 stimuli from dataset 2. In S2 Fig, we show the confusion matrices for all the remaining data.

Fig 5. Pipeline for semi-supervised clustering of visual areas.

A) Block diagram of the clustering steps. B) Initial clusters that are labeled using the retinotopic map. C) Neighbors of a labeled cluster for area AM. BIC score is computed between these neighbors, and closest few are merged every iteration. D) An intermediate step in the clustering process. E) Final clustering result with accuracy (ACC).

Fig 6. Boundaries obtained by clustering visual cortex areas using the semi-supervised pipeline.

A) Boundaries derived using different visual stimuli in one mouse. Visual stimuli and accuracy (in %) are noted. B) Boundaries obtained for different mice using natural movies as stimuli. C) Boundaries obtained for different initializations of UBM for the same mouse and keeping visual stimuli unchanged. D) Boundaries derived from resting state responses.

Fig 7. Accuracies obtained by the supervised/semi-supervised pipeline with varying response lengths of resting state and stimulus-induced responses.

A, B) Results for Mouse M4 and M5, respectively, using the supervised approach. C, D) Results using the semi-supervised approach. E, F) Results for the two-photon dataset, using the Cre-lines Emx1-IRES and Nr5a1, respectively.

Fig 8. Intra-area and inter-area correlations computed on input responses and LDA features.

A–D) Correlations computed from mouse M4 of wide-field dataset. E–H) Correlations computed from Emx1-IRES Cre-line of two-photon dataset. The correlations are computed as averages over all unique pairs of neurons/pixels in the test data, which were not used to train the LDA projection matrix. The correlations in the two-photon dataset are computed using data pooled from different mice and multiple sessions. In S3 and S4 Figs, we present a detailed correlation analysis with information about individual mice and session. Further examples of correlations analysis are shown in S5–S7 Figs.

Fig 9. Two-dimensional representation of the supervised LDA subspace.

A, B) LDA subspace of wide-field dataset (mouse M4). C, D) LDA subspace of two-photon dataset (Cre-line Emx1-IRES). The plots on the left (A, C) are obtained from natural movie responses and that on the right (B, D) are obtained from resting state responses.