Fig 1.
PUPAID is able to process raw immunofluorescence data, extract cell-based information and convert them into TSV- and XLSX-formatted tables as well as into cytometry FCS files for easy and thorough analysis.
Table 1.
Metrics used or produced by PUPAID.
This table summarizes all the metrics which are used or produced through PUPAID workflow, as well as their full names, formulas and short descriptions. All metrics except CTCF are directly computed and exported using the Measure feature in ImageJ. CTCF for each parameter are computed afterwards via R during the generation of the merged FCS file using the formula presented in the table.
Fig 2.
Local contrast enhancing using CLAHE algorithm is necessary for optimized cell segmentation by PUPAID.
(A) Stitched image from the test dataset which is overlaid with the 10 200x200 pixels crops we generated to evaluate PUPAID’s performance. 5 of them are located in high-density regions (red squares) and the remaining 5 are located in low-density regions (yellow squares). Then, we applied on these crops the methodology described in the Benchmarking of the different cell segmentation approaches subsection of the Material and Methods section of the manuscript. (B) Percentage of detected objects for each depicted method in HD (left) and LD (right) regions as compared to the manual cell segmentation which represents 100% of the cells. (C-D) Graphs showing the Intersection over Union (IoU) thresholds vs. the 4 quality metrics (Precision, Recall/sensitivity, Jaccard and F-measure) generated by the MiC ImageJ plugin for the depicted methods in HD (C) and LD (D) regions. (E) Same data as previously shown in (C-D) except that this subfigure focuses on an IoU threshold of 0.4 for HD (top) and LD (bottom) regions. Groups were compared using Student’s t-test and p-values were reported as follows: *: p-value < 0.05, **: p-value < 0.01, ***: p-value < 0.001 and ****: p-value < 0.0001. Data are represented as mean±SEM.
Fig 3.
Benchmarking of either PUPAID-, manual- or other state-of-the-art methods-generated cell segmentation.
(A) Original greyscale images and associated cell segmentation masks generated with the depicted methods from 1 representative high-density (HD) and 1 representative low-density (LD) crop coming from the test dataset. We extracted 10 distinct crops of dimensions 200x200 pixels from the test dataset: 5 coming from HD regions and 5 coming from LD, as respectively shown by the red and yellow squares in the Fig 3A. Then, we applied on these crops the methodology described in the Benchmarking of the different cell segmentation approaches subsection of the Material and methods section. (B) Percentage of detected objects for each depicted method in HD (top) and LD (bottom) regions as compared to the manual cell segmentation which represents 100% of the cells. (C) Graphs showing the Intersection over Union (IoU) thresholds vs. the 4 quality metrics (Precision, Recall/sensitivity, Jaccard and F-measure) generated by the MiC ImageJ plugin (accessible here: https://github.com/MultimodalImagingCenter/MiC) for the depicted methods in HD (left) and LD (right) regions. (D) Same data as previously shown in (C) except that this subfigure focuses on an IoU threshold of 0.4 for HD (left) and LD (right) regions. Groups were compared using Student’s t-test and p-values were reported as follows: *: p-value < 0.05, **: p-value < 0.01, ***: p-value < 0.001 and ****: p-value < 0.0001. Data are represented as mean±SEM.
Table 2.
Summary of the efficacies of the compared cell segmentation methods.
The table presents the mean values independently for each measurement, which include the percentage of detected objects and the Precision, Recall/sensitivity, Jaccard and F-measure metrics generated by the MiC ImageJ plugin. The presented values were computed with an IoU threshold of 0.4 and with the 1:1.4 σ ratio for PUPAID. The efficacy of a given method is reported as a percentage between 0 and 1 which is relative to the manual cell segmentation.
Fig 4.
Gating strategy-based analytic approach on the produced FCS files.
The final goal of PUPAID is to produce FCS files which can be conveniently opened and analyzed in conventional flow/mass cytometry software such as BD FACSDiva (BD Biosciences), FlowJo (BD Life Sciences), CytExpert (Beckman Coulter), CyTOF (Fluidigm), Kaluza (Beckman Coulter), or FCSExpress (De Novo Software) or through dedicated R-written all-in-one analysis pipelines such as PICAFlow [34], cytofkit [35] or CyTOF [36]. Each plot can represent a given dimension of the data generated during PUPAID processing and analysis, and each dimension represents an aspect of each cell that was segmented during the analysis, such as areas, X/Y coordinates, shape descriptors or fluorescence markers. Red circles indicate examples of gates that can be drawn in order to extract the cells of interest.