Fig. 1
From: CenFind: a deep-learning pipeline for efficient centriole detection in microscopy datasets
Markers used to detect centrioles, procentrioles and PCM. a Schematic representation of centriole (blue) and procentriole (red), with locations of centriole, procentriole and PCM markers used in the dataset. b Schematic representation of centriole duplication cycle. M: Mitosis; G1 Gap 1 phase; S: synthesis; G2: Gap 2 phase. c. Representative image of human hTERT-RPE-1 cells immunostained for centrin (red), Cep63 (green), PCNT (cyan) and stained for DNA (blue); box indicates region magnified in d–g. d–g Individual channels: Centrin (e), Pericentrin (f) and Cep63 (g). Arrows indicate foci positions. Note the difference in shape and intensity between centriolar markers. h The input for CenFind is a multi-channel z-stack, which is max-projected along the z-axis. Channels of interest are selected and passed onto the respective models for centriole/procentriole/PCM detection with SpotNet (left) and nucleus segmentation with StarDist (right), before saving as coordinates (left) or masks (right). For cell scoring, the nuclei are assigned to centrioles/procentrioles/PCM using a distance metric; the analysis can be carried out using the original image data, together with the saved positions of centrioles/procentriole/PCM. i–m F1-scores on the test set at different pixel (bottom X axis) and distance (top X axis) tolerances, defined as the distance between an annotation and a prediction. The model needs to yield excellent performance for tolerances as low as 3 pixels to tell apart centriole and procentriole, which are a mere ~ 300 nm away from one another