Movie segmentation: area-finding

Processing fUS data is greatly enabled by identifying coactive components in the video, much like widefield data and fMRI. For this purpose we leverage recent graph-based approached from the widefield calcium imaging literature, in particular the recent Graph-Filtered Time-trace (GraFT) analysis. GraFT estimates a graph between pixels (each pixel is one node) and connects the nodes based on the temporal correlation between pixels. Once the graph is obtained, it is used to bias a dictionary learning procedure that learns a set of component time-traces that can sparsely reproduce each pixel on the graph.

We find the GraFT can identify, with no a-priori knowledge of where in the image each pixel is, coherent groups of pixels that are spatially colocalized. Interestingly we observe both components completely interior to, as well as straddling the boundary between, anatomical boundaries identified by aligning to a standard atlas. In addition to exposing the spatial structure of correlated activity, GraFT components also provide an interpretable dimensionality reduction where each video frame can be represented only by the activity level of each component (a much smaller number of values). This dimensionality reduction further eases ensuing analyses, such as behavioral modeling.