Search

Additional file 3: Figure S3. Ciliary FP/FN rates across separate experiments or different Ground Truths within the same experiment. (A) These are the analyzed images of RPE cells with Arl13b-stained cilia and DAPI-stained nuclei from two separate experiments (Experiment 1 and Experiment 2). The images for Experiment 1 were the same images that were analyzed for the RPE Arl13b cilia condition in Fig. 6b. (B) The images from Experiment 2 were analyzed two times using two different representative images to establish two different Ground Truths. The two different representative images from the Experiment 2 data set are shown, and below are their respective parameter threshold values after detection auto-optimization. (C) Manual and automated detection of RPE primary cilia labeled with Arl13b in 60×-mag. images. Manual true candidates (TC) represent the total number of true cilia (light green bar). Automated TC represents the number of manual TC that was detected in automated analysis mode (dark green bar). Across different experiments (i.e. Experiment 1 vs. Experiment 2), the FP rates, FN rates, and F1 scores were similar. When repeating the same experiment (i.e. Experiment 2), but using different representative images of the data set for detection auto-optimization, FP and FN rates were similar (1–3%) and F1 scores were identical (F1 = 0.98). (D) Tabular data for the values depicted in (C). Accuracy ratings are based on F1 score values, which range from 0 to 1.00 (0.95–1.00, ++++ ; 0.90–0.94, +++; 0.85–0.89, ++/+++). All measurements are reported as averages ± standard deviations of multiple images from one experiment.

Additional file 6: Table S6. Tabular data of Fig. 7. Images of RPE cells with Arl13b-stained cilia, taken at two different magnifications, were analyzed manually, automatedly, and semi-automaticallywith the ACDC software. Analysis times were recorded for each image and then averaged. Analysis times for fully automated analysis (“Auto) are greatly faster than those of manual analysis. Standard deviations of fully automated analysis times are much smaller than those of semi-automated analysis times (“Semi-Auto”), which are smaller than those of manual analysis times. Analysis times include the time required to count nuclei, count cilia, and measure cilia length. All measurements are reported as averages ± standard deviations of multiple images from one experiment.

Additional file 4: Figure S4. Quantitative analysis of ACDC accuracy in detecting nuclei and cilia frequency. (A) Examples of nuclei detection with ACDC software of the two representative images used for Experiment 2. (B) Manual and automated detection of RPE nuclei labeled with DAPI in 60×-mag. images. Manually-detected nuclei represent the total number of true nuclei (light green bar). Automatedly-detected nuclei represent the proportion of true nuclei that were detected in automated analysis mode (dark green bar). In regards to nuclei detection, repeating the analysis using two different representative images from the same Experiment 2 data set yielded identical FP and FN rates (0.9%) and identical F1 scores (F1 = 0.99). Note that detection auto-optimization from representative images is not utilized for nuclei segmentation and detection. (C) Tabular data for the values depicted in (B). Accuracy ratings are based on F1 score values, which range from 0 to 1.00 (0.95–1.00, ++++ ; 0.90–0.94, +++; 0.85–0.89, ++/+++). (D) Manual and automated analysis of cilia frequency for the images from Experiment 2. Cilia frequency values were calculated by dividing the cilia count/image values by the nuclei count/image values. All measurements are reported as averages ± standard deviations of multiple images from one experiment.

Additional file 1: Figure S1. Surface plots of F1 scores for different cilia reporters. Although the F1 score is ultimately based on four parameters, here we depict a 3D surface plot with two of the parameters on the x- and y-axes and F1 score on the z-axis for visualization purposes. Each plot is based on the corresponding image of ciliated RPE cells. F1 score values for automated analysis of each image were calculated by incrementally increasing one parameter threshold while holding the other parameter threshold constant and then counting the number of FPs and FNs. These plots illustrate that there could be many parameter combinations that results in the same maximal F1 score, in which case the program’s default is to choose the parameter combination that minimizes all parameter values while retaining that maximal F1 score. Of the three cilia reporters, Arl13b seems to be the ‘best’ reporter, as its surface plot has the largest area at F1 = 1.00. Thus, at most combinations of SNR and STD threshold values, all Arl13b-labeled true cilia will be detected and no FPs will be included.

Additional file 5: Figure S5. High cell confluency can affect automated cilia and nuclei detection and cell frequency measurements. (A) Analyzed images from two separate samples of differing cell confluency (low and high). (B and C) Bar graphs and corresponding tabular data of manual and automated analysis of RPE Arl13b-labeled cilia in 60×-mag. images. Manual true candidates (TC) represent the total number of true cilia (light green bar). Automated TC represents the number of manual TC that was detected in automated analysis mode (dark green bar). Despite differences in cell confluency, FP and FN rates for cilia detection remained between 1 and 4% and F1 scores were similar (low confluency, F1 = 0.98 vs. high confluency, F1 = 0.97). (D and E) Bar graphs and corresponding tabular data of manual and automated analysis of RPE nuclei in the same 60×-mag. images. FP rates for nuclei detection were greater in low confluency images than in high confluency images (4.5% vs. 0%), but FN rates for nuclei detection were much greater in high confluency images than in low confluency images (12% vs. 1%). Consequently, the large FN rate of 12% for nuclei detection in high confluency images yielded a worse accuracy score (F1 = 0.94) compared to that of low confluency images (F1 = 0.97). (F) Manual and automated analysis of cilia frequency for the low and high confluency data sets. Cilia frequency was calculated by dividing the cilia count per image values by the nuclei count per image values. Cilia frequency remained nearly identical between manual and automated analysis for low confluency images, but cilia frequency values varied more between manual and automated analysis for high confluency images (manual, 39% ± 12% vs. automated, 45% ± 17%). This difference is due to the disparity in nuclei detection between manual and automated analysis for high confluency images. All measurements are reported as averages ± standard deviations of multiple images from one experiment.

Additional file 2: Table S2. Tabular data of Fig. 6. For each group (e.g. AcTub-labeled cilia in NIH3T3 cells, etc.), absolute cilia count from analyzed images were averaged and reported as manual true cilia (“Manual TC”) averages ± standard deviations. Thus, a data set with a Manual TC of 11.9 ± 2.8 indicates an average of 11.9 cilia per image with a standard deviation of 2.8 cilia per image. “Automated TC” average counts differ from those of “Manual TC” because in automated analysis mode some true cilia might not be detected by the software. An “Automated TC” of 11.4 ± 2.4 (96% of “Manual TC”) indicates that, on average, the software auto-detects 96% of true cilia in each image. A false positive (FP) rate of 2.7 ± 1.9 (23% of “Manual TC”) indicates that, on average, the software will include 2.7 additional objects (23% more fake cilia) in each image, on top of the 11.4 ± 2.4 (96%) detected true cilia per image. A false negative (FN) rate of 0.5 ± 0.9 (4% of “Manual TC”) indicates that, on average, the software will not detect 4% of the “Manual TC” in each image. Precision (α) is calculated using the data from the “Automated TC” and “FP” rows. Recall (β) is calculated using the data from the “Automated TC” and “FN” rows. F1 score is calculated using the data from the “Precision (α)” and “Recall (β)” rows. Accuracy ratings are based on F1 score values, which range from 0 to 1.00 (0.95–1.00, ++++ ; 0.90–0.94, +++; 0.85–0.89, ++/+++).

aaaaaa