Figure 5

Comparison of spot identification and classification methods. A. The upper left is a maximum merge projection of an smFISH image from a C. elegans embryo for which 488 signal spots were counted by hand (using the GUIs described here). A green rectangle highlights a section of the image. The other three images show spots identified in this green rectangle by FISH-Quant (upper right), the threshold-picking method (lower right), and the method described here (Aro: lower left). The number of signal spots identified in the embryo by the various methods are noted in the lower right of each image. Circles mark the locations of identified spots and are color coded by z-slice. Arrows point to representative areas depicting the tendency of threshold method to identify a large high intensity region comprised of several spots as a single spot. B. A plot of manually counted spot number (x-axis) and estimated spot number (y-axis) by Aro, threshold-picking, and FISH-Quant across 28 C. elegans embryos. Both FISH-Quant and threshold-picking tend to underestimate the true number of spots (particularly at higher spot counts for the threshold method) while our Aro machine learning method performs well across a range of spots numbers. Spearman correlations (r) between the true and estimated spot number are listed for each method. Both Aro and threshold-picking perform significantly better than random on this dataset. Interval estimates are depicted for Aro. Neither FISH-Quant nor threshold-picking provides a way to estimate error.