Automatic identification of fluorescently labeled brain cells for rapid functional imaging

Abstract

The on-line identification of labeled cells and vessels is a rate-limiting step in scanning microscopy. We use supervised learning to formulate an algorithm that rapidly and automatically tags fluorescently labeled somata in full-field images of cortex and constructs an optimized scan path through these cells. A single classifier works across multiple subjects, regions of the cortex of similar depth, and different magnification and contrast levels without the need to retrain the algorithm. Retraining only has to be performed when the morphological properties of the cells change significantly. In conjunction with two-photon laser scanning microscopy and bulk-labeling of cells in layers 2/3 of rat parietal cortex with a calcium indicator, we can automatically identify ∼ 50 cells within 1 min and sample them at ∼ 100 Hz with a signal-to-noise ratio of ∼ 10.

Fig. 1.

Examples of annotations used to generate the 2 classifiers. A: example, shown raw and annotated, for the 1st classifier. This image is 1 of 16 full field images annotated in Adobe Photoshop. Because 4 trials were performed for each annotated region, these annotations were used in learning on 64 stimulation trials. Green indicates that the pixel is part of cell somata, whereas blue indicates that it is uncertain whether a pixel is part of cell somata. All uncolored pixels are taken as examples of pixels that are not parts of cells. Notice that a very rough annotation was sufficient to produce good results. B: example to generate second classifier. A screenshot of a graphical user interface used to annotate whether a particular cluster of pixels is a cell, not a cell, or ambiguous region. Left: a large mean image is shown with a current candidate cell outlined. Outlines that have not yet been evaluated are colored blue, those that were selected as not cells are colored red, those that have been selected as cells are colored green, and those that were selected as ambiguous regions remain colored blue. Top right: a normalized mean image of the region. Bottom right: a mean image with all of the previously made selections outlined with appropriate colors.

Fig. 2.

Example of segmentation of a test data set. A: the 4 unnormalized filtered version of the raw data (Table 1, formulas 1–4). The color corresponds to amplitude of the filtered output. The normalized versions of filtered images from A (Table 1, formulas 5–8). B: the output of the 1st classifier. The color corresponds to the likelihood that a given pixel is a cell. C: the output of the 2nd step classifier, with isolated pixels, i.e., speckle noise, removed with a 5 × 5 pixel median filter, along with the output values thresholded to form clusters of pixels that are candidate cells; we chose 6 levels, which correspond to pixels lying in the top 5, 10, 15, 20, 25, and 30% of the maximum amplitude. D: final classification made by thresholding the output shown in C.

Fig. 3.

Validation statistics for the classifiers. A: a histogram of cross-validated examples binned by the scores they have received from the first classifier. Black are examples of pixels that are parts of cells, whereas gray are examples of pixels that are not parts of cells. B: receiver operating characteristic curve of the 1st classifier; the 2 dotted lines indicate the point on the ROC curve for which the score threshold is zero. Note that because ground truth is poorly defined, the ROC curve is only approximately representative of the real classifier errors. C: a histogram of cross-validated examples binned by the score they have received from the 2nd classifier. Black are examples of candidate cells that are actually cells, whereas gray are candidate cells that are not cells. D: an ROC curve of the 2nd classifier; the 2 dotted lines indicate the point on the ROC for which the score threshold is zero, which is the nominal final threshold for our algorithm. Note that because ground truth is poorly defined, the ROC curve is only approximately representative of the real classifier errors.

Fig. 4.

Two examples of cell segmentation and fast scanning for functional imaging of neurons and astrocytes in rat parietal cortex. A: a full-field image of a region with 68 cells, obtained at 4 frames/s, with a scan path superimposed on it in which all cells are sampled at 70 Hz. The green channel shows the fluorescence from Oregon Green Bapta-1, whereas the red channel shows fluorescence from Sulforhodamine 101. White shows the outlines of cells as determined by our algorithm. B: part of the raw data output from consecutive scans, including a hindlimb stimulation. C: activity of 10 cells, 9 neurons, and 1 astrocyte as indicated in A and B, during the same time interval as shown in B. The traces shown in the order of the cells that were scanned and represent typical results. D: distribution of onset times for changes in intracellular [Ca²⁺] in all 68 cells after stimulation across 9 trials. E: a full-field image of 19 neurons, 1 astrocyte, and 3 blood vessels scanned at 110 Hz with a scan path superimposed on it. F: part of raw data output that includes a hindlimb stimulation event. G: activity of cells, neurons, and an astrocyte indicated in E and F during the same time interval as shown in F. H: the calcium response of the astrocyte (A1), the average neuronal response (N1–N19), and the speed of red blood cells in one capillary (V1).