Fig. 2: FAST enhances Ca²⁺ imaging quality and analytical accuracy.

a Raw GCaMP6s calcium imaging videos from the mouse vS1 region were first denoised using different denoising methods. Subsequently, temporal maximum intensity projection (MIP) was applied to the videos. The resulting projection images were then segmented using Cellpose. b In vivo neural population calcium imaging data were obtained from ref. ³² The top panel shows the raw calcium imaging data with a red box indicating a region of interest (ROI). Enlarged images of the ROI at four different time points during a single spike event (~7.5 s time window) are presented, illustrating neuronal activity in various states to evaluate denoising performance. The bottom panel displays the corresponding segmentation results, with manual annotations used as ground truth: correctly segmented regions (true positives) are shown in green, missed regions (false negatives) in red, and extra regions (false positives) in blue. Three neuron regions are highlighted with boxes in the bottom panel, and their corresponding magnified segmentation results are shown in (h). Scale bar, 100 μm. c–g Example denoising and segmentation results from raw data processed using DeepCAD-RT, SRDTrans, DeepVid, SUPPORT, and FAST. Following the same structure as (b). In g, arrowheads indicate the fine dendrites that become observable after FAST denoising. The images shown are representative frames, and similar results were obtained across all 8000 frames analyzed. h Example neurons, zoomed in from the boxed regions in (b–g), illustrating segmentation results across different denoising methods. Each row represents a single neuron, while each column shows the maximum intensity projection of the raw data or data denoised by one of the five methods. To ensure a fair visual comparison, all magnified views shown were cropped from full-sized images that had been globally normalized to a single, unified intensity range, with no subsequent, individual re-normalization applied. In h, asterisks indicate neurons that were successfully segmented by Cellpose. Scale bar, 5 μm. i Segmentation performance (Accuracy, Recall, and F1 score) was calculated for n = 155 neurons, all located within a single imaging dataset.