In the current study, an automated cell counting algorithm was created to quantify resultant migratory 4T1 breast cancer cells in cell migration assays that used 10,000 and 100,000 total seeded cells. The algorithm was only tested in migration assays conducted in modified Boyden chambers using DAPI-stained 4T1 breast cancer cells. For a total of 47 images, automated cell counts had strong correlation with manual counts (r2 = 0.99, P < 0.0001; Figure ). To highlight the congruency between methods, the y-intercept for the linear regression line (Figure ) indicated that our automated method overestimated an average of only 2 cells under all experimental and imaging conditions. Furthermore, there were no differences in cell counts between methods regardless of experimental conditions (number of cells seeded) or objective power (Figure ).
Our results support the notion that manual cell counting is time consuming (47 images; manual counting time = 2.4 hours versus automated computing time = 14.5 seconds) and subject to operator bias (hence our experimenter was blinded). Further, it is difficult to reproduce exact measurements using manual methods (hence variability in our manual cell counts, Figure ), which is likely due to disparate criteria under which manual cell counting is performed from image to image. Our automated method processes and counts all images using set criteria making it immune to the aforementioned sources of error. Upon multiple independent screening of the same 47 images, zero variability was still obtained.
Quantification of migrated cells from cell migration experiments are generally limited to objective powers ranging from 5 × to 200 × [14
] and thus require 3 [15
] to 10 [16
] fields of view in an effort to capture a representative sample of the membrane. The consistency in computed cell counts among different objective magnifications highlights our algorithm's ability to accurately resolve and count cells even under low magnification; thus enabling the user to analyze larger proportions of total membrane surface area in one field of view. This is due to the thresholding portion of our algorithm, which optimizes contrast between cell nuclei and the image background (Figure ).
Using our algorithm, the number of cells counted depends on the surface area of the majority of nuclei in the image. Therefore, if the majority of cells are multinucleated, then the representative (mean) nucleus area will be greater than if the majority of the cells had single nuclei. We describe (Results - Cell Counting Algorithm section) that our algorithm uses the Matlab® command round by which we account for discrepancies in nucleus sizes based on ratios (discrete nucleus surface area/representative nucleus surface area). Under this command, nuclei area ratios of less than 0.5 are discarded and not counted (this accounts for image artifacts, cells that have not completely migrated through pores, cellular debris, etc). Ratios including 1.0 and between 1.0 and 1.5 are counted as single cells, and those including 1.5 and between 1.5 and 2.0 are counted as 2 cells (and so on). Thus, using this ratio and rounding technique our algorithm accounts for biological variations in nuclei size (within a reasonable range). This mathematical approach is definitive, reproducible, and invariable, which is in contrast to the gold standard of manual counting that is dictated by "experience" and "artistic impression".
Irregularities in nuclei shape do not affect how the algorithm calculates average cell nucleus area, as it does not constrain expressed fluorescent area to a specific shape. In other words, assuming surface area is maintained across the majority of cells (as addressed above), fluorescent cell nucleic area is assumed to be independent of shape.
In the current study, using 4T1 cancer cells, we did not observe an abundance of atypical nucleus sizes or shapes. Our research addressed the issue of overlapping cells by using increasing numbers of cells seeded into Boyden chambers (from 10,000 to 100,000 cells). This was based on the reasonable postulate that increased cell number would decrease the distance between migrated cells, thus increasing the opportunity for migrated cells to overlap and be miscounted. We saw no difference in average cell area (mean ± S.E.M.) regardless of total number of seeded cells.
To substantiate the accuracy of our approach, there were no differences in the mean number of cells counted by our algorithm versus those manually counted by a blinded experienced cell biologist (the gold standard). However, repeated manual counts produced intra-observer variability, whereas the counts by our algorithm were devoid of variability.