An automatic method for robust and fast cell detection in bright field images from high-throughput microscopy

BMC Bioinformatics

Table 1 Manual evaluation of segmentation results

Experiment time	12 h	2 d	4.5 d	5.5 d
Number of manually counted cells	37	90	5837	7414
Correct	34(91.9%)	74(82.2%)	5356(91.8%)	6638(89.5%)
Missed	1(2.7%)	4(4.4%)	78(1.3%)	86(1.2%)
Over-segmented	2(5.4%)	10(11.1%)	230(3.9%)	411(5.5%)
Under-segmented	0(0.0%)	0(0.0%)	46(0.8%)	87(1.2%)
Debris	6	7	1	32
True positives	36	84	5632	7136
False positives	6	13	96	329
False negatives	1	6	205	278
Accuracy	0.83±0.11	0.82±0.03	0.95±0.02	0.92±0.03
Specificity	0.86±0.11	0.87±0.06	0.98±0.01	0.96±0.03
Sensitivity	0.97±0.06	0.94±0.04	0.96±0.02	0.96±0.01

Two randomly chosen fields of view per well were quantified for 12 hours, 2 days, 4.5 days and 5.5 days, respectively. In each field of view, the number of true cells was counted. All segmented objects were classified as correct, over-, under-segmented, or debris. Accuracy, Sensitivity and Specificity of cell detection were calculated based on true positives (complete cell bodies, the largest fragment of over-segmented cells and one cell per under-segmented object), false positives (dirt and cell fragments) as well as false negatives (missed cells, cells in under-segmented objects). Note that we deliberately keep differences in the total number of counted cells at different experiment times, since these impact on the standard deviation of accuracy, specificity and sensitivity.

ISSN: 1471-2105