CenFind: a deep-learning pipeline for efficient centriole detection in microscopy datasets

BMC Bioinformatics

Table 2 Model comparisons

Method	F₁-score on test set (mean ± st. dev.)
CenFind (Multiscale U-Net)	0.908 ± 0.043
U-Net	0.868 ± 0.073
Laplacian of Gaussian (LoG)	0.81 ± 0.230
OpenCV's Simple Blob Detector	0.615 ± 0.199
FociDetector	0.586 ± 0.254

Comparison of F₁-scores (mean ± standard deviation) obtained on the test set (N = 15 fields of view) using the model of CenFind (Multiscale U-Net), the same model ablated in its multiscale (U-Net), the Laplacian of Gaussians (LoG), the OpenCV Simple Blob Detector, the FociDetector model from (4). Note that FociDetector uses a series of convolutions to highlight regions likely to have centrosomes and multiply them with the input, to compensate for the loss in resolution generated by the convolutions. Overall this results in a relatively high precision (0.852 ± 0.165), but in a substantially lower recall (0.519 ± 0.311)

ISSN: 1471-2105