Fig 1.
Image of CMOS-based device and implantation schematic.
(a) Photograph of implantable CMOS-based imaging device. (b) Coronal (-4.7 mm) and (c) sagittal (0.12 mm) brain slice schematic of device implantation into the DRN.
Fig 2.
Overview of the ROI determination algorithm.
A specified number of frames are adaptively binarized and then cleaned with a sequence of morphological opening and morphological area opening. The images are averaged, binarized again and cleaned with the same sequence of morphological processes.
Table 1.
Coefficient of each parameter after assessing partial least squares regression to average ROI area.
Fig 3.
Comparison of ROI masks generated using different algorithm parameters.
(a) Average image after 1, 10, 20, 30, 40, 50 and 60 minutes of Ca2+ imaging. (b-d) ROI mask generated after running the algorithm on the corresponding Ca2+ imaging data when the parameters are: (b) σ1 = 1.5, FP1 = 2, A1 = 0, σ2 = 0.5, FP2 = 3, A2 = 0, (c) σ1 = 2.5, FP1 = 2, A1 = 0, σ2 = 2.5, FP2 = 2, A2 = 0, and (d) σ1 = 1.5, FP1 = 3, A1 = 0, σ2 = 1.5, FP2 = 3, A2 = 0. The first set of images (a) illustrates a typical case of too few ROIs. The second set (b) exhibits a typical case of too big ROIs. The last set of images (c) exhibits uniformly distributed and sized ROIs.
Fig 4.
Comparison of ground truth position of simulated neurons with different firing patterns and ROI selection by algorithm.
(a) The position of each simulated neuron in the simulated field of view is displayed with its corresponding activity pattern. (b) Pattern 1 neurons exhibit high activity in the first 10 minutes, medium activity in the next 20 minutes, and low activity in the last 60 minutes. In pattern 2 neurons, the high activity is shifted to the second 10 minutes and the first 10 minutes is medium activity. Unresponsive neurons maintain baseline activity throughout the simulation.
Fig 5.
Clustering of ROIs and average trace of each ROI cluster.
(a) Ground truth position of simulated neurons with ROIs selected by the algorithm and clustered with PCA and k-means. (b) Average ΔF/F0 of clusters 1 to 3 which closely resembles the three ground truth activity patterns.
Fig 6.
Accuracy of ROI algorithm in detecting simulated neuron position over 1000 simulations.
For each simulation, the number of neurons tagged as ROIs was counted and divided by the total number of neurons of that activity pattern. This histogram shows that the neurons that are not detected are the inactive neurons. Over 400 simulations correctly marked 100% of the active neurons.
Fig 7.
DRN activity recorded with CMOS-based device after formalin injection.
(a) Average DRN fluorescence frame every 5 minutes shows clear difference between prestimulation and poststimulation frame average. (b) Average fluorescence time series of each ROI shows that difference in activity among ROIs. (c) Difference between %ΔF/F0 traces with and without spectral subtraction.
Fig 8.
Average trace of each ROI cluster at maximum silhouette score.
The ROIs were clustered with PCA and k-means to functionally differentiate the ROIs between active and non-active. Clear differences between the activity patterns of each cluster can be observed proving the effectiveness of the grouping.
Fig 9.
Groupwise comparison of formalin- and PBS-injected mice’s DRN activity using ROI algorithm and whole frame averages.
Average trace of highest activity ROI clusters (a) of each mouse in each treatment group compared to the average of the whole image (b). The average intensity of the formalin group is higher than PBS group 5 minutes after the stimulation and even throughout the 60-minute experiment. This is not the case when analyzing the data using only whole image averages.