Ventral and Dorsal Stream Dissociation During Action Recognition in the Human Brain.

Introduction. According to the perception-action framework [4], the ventral stream mediates object recognition, while the dorsal stream processes the sensory control of object-directed actions. The existence of two functionally independent visual streams [3], however, is debated [8]. To determine whether such a pure functional dissociation exists, or rather an integration of action-object representations occurs in the ventral and dorsal streams, we used a Multi Voxel Pattern Analysis to examine whether neural patterns of visually perceived hand-mediated actions differed on the basis of the object class being manipulated. Methods. We used an fMRI (Philips 3T; TR 2.5s; 3x3x3 mm voxel size) six-run slow event-related design to examine neural activity in 14 right-handed healthy volunteers while they watched 3s-long videos that randomly alternated between different types of human hand-made, object-directed actions or environmental scenes with 7s of ISI. A first set of stimuli depicted ‘tool-mediated’ (e.g. hammering a nail, cutting with scissors), ‘intransitive’ (e.g. clapping, ok gesture) and ‘distal transitive’ (e.g. opening a jar, grabbing) actions. A second set consisted of videos reproducing three fixed motor acts (‘pushing away’, ‘grasping-to-lift’, ‘putting down’) performed with ‘animate’ and ‘inanimate’ objects. After standard processing using AFNI [1], BOLD responses to each stimulus of the first set were used in a 3-class Support Vector Machine (SVM) classifier [6] with an additional Recursive Feature Elimination algorithm [2]. Then, the discriminative voxels of this classifier responding to action features were isolated into ventral occipito-temporal [5] and dorsal parietal regions of interest (ROIs). Inside these ROIs, the patterns of BOLD response elicited by the second set of action stimuli were compared to each other and to those elicited by environmental scenes to obtain a representational dissimilarity matrix (RDM) using the 1-r Pearson coefficient [7]. A hierarchical clustering procedure was derived from the RDM to create a dendrogram. Moreover, we used multi-class SVM classifiers to assess the uncertainty in cluster analysis. Accuracy (Acc) values of the classifiers were tested as significantly different from chance by a permutation test. Results. The SVM classifier, trained on the first set of stimuli, was able to distinguish between ‘tool-mediated’, ‘intransitive’ and ‘distal transitive’ actions with an Acc of 73.3% (p<0.0025). The discriminative map [Fig 1] relied on an inferior frontal, premotor, inferior parietal and temporal cortical network, and included brain areas mainly within the human mirror system [9]. Using the second set of actions, as expected, environmental scenes were clearly separated from actions in both ROIs (Acc 92%, p<0.0025). Within the inferior temporal ROI, the RDM showed a significant distinction between the two classes of objects only (‘animate’ vs. ‘inanimate’, Acc 69.7%, p<0.0025) but was unable to recognize the different actions [Fig 2]. Using the parietal ROI, we found a significant distinction between three classes of actions (‘pushing’ vs. ‘grasping’ vs. ‘putting’ acts, Acc 39.8%, p<0.05) and between the two classes of objects (‘animate’ vs. ‘inanimate’, Acc 58.3%, p