Is there any way this plugin could be adapted to compute the average precision or mean average precision? I don't know how to code and cannot do it myself. Thank you.

[gopher-confocal]

[cmessaoudi]

Hi,
It is certainly possible to add new metrics.

I do have a question regarding the average precision (AP) metric you requested: there are several definitions in use. If you are referring to the definition used by Cellpose and in some light microscopy papers, it is actually equivalent to the Jaccard index as provided by this plugin, calculated as:
AP = TP / (TP + FP + FN)

Once you confirm which definition you prefer, I will update the implementation for AP and mAP accordingly.

Best,

[gopher-confocal]

My understanding is that the average precision is the area under the precision vs recall curve and the mean average precision is the average of the APs from multiple classes. I'm going from the definitions of AP and mAP in the text on page 2 of this paper: [https://www.nature.com/articles/s41592-023-01942-8].

Thank you for responding and working on this! I appreciate this tremendously!

[cmessaoudi]

In the paper, the authors note that several definitions of Average Precision (AP) exist in the literature. The definition relevant to your needs is the first one listed in their table, while the plugin currently implements the second definition (also known as the Jaccard index).
I will implement the first definition as soon as possible.
Regarding the mean Average Precision (mAP), it is calculated as the average of AP across all classes. At present, the plugin is designed for instance segmentation without class differentiation, but extending it to support class-specific segmentation is both feasible and straightforward.
One remaining question concerns the input data format. I propose using a multi-channel image, where each channel represents the instance segmentation for a specific class.
Would this format meet your requirements?

[gopher-confocal]

Sure, a multichannel format works for me. Would the plugin still have the option for a single class of label objects?

I really appreciate this enormously! Thanks a million! I wish there were some way I could reciprocate.

[gopher-confocal]

Hello! I request an update on the progression of the plugin update that I asked for, please. How far along have gone regarding the average precision calculation, using the definition of the AP as the area under the precision-recall curve? Thank you!

[cmessaoudi]

Apologies for the delay—I’ve been juggling several priorities and only had less than an hour last week to dedicate to this project.
By the way, do you have access to the confidence scores for each of your detected objects? If so, how do you plan to pass these values to the plugin? In my view, the best approach would be to use a CSV file, with one line per object and in the same order as the mask.
Currently, the plugin doesn’t support this type of input, which is necessary for generating a precision-recall curve. Modifying the code and GUI to accommodate this will take some time. If you don’t have access to these confidence values (for example, Cellpose doesn’t provide them), you’ll need to rely on the Jaccard index instead.

[gopher-confocal]

I don't know how to get the confidence scores for my segmented cells. How do I do that, please?

[cmessaoudi]

Which segmentation or detection algorithm are you currently evaluating? For example:

Segmentation-focused: Cellpose, StarDist, U-Net (or variants)
Detection-focused: R-CNN, YOLO, DTER, SAM

To clarify: U-Net–based approaches (including Cellpose and StarDist) typically do not output per-object confidence scores, while object detection algorithms—whether standalone or combined with segmentation—usually do.
You’re welcome to message me directly at my email address, which is publicly available online.

[gopher-confocal]

I'm comparing various open-source software for segmentation benchmarking: Cellpose3 & 4, napari-CellSAM, QuPath & StarDist plugin, CellProfiler & StarDist plugin.

…

________________________________________________ Mary E. Brown, Pharm.D. Imaging Specialist/Research Professional 3/Lead Image Data Analyst University Imaging Centers Snyder Hall, rm. 23F 1475 Gortner Avenue Saint Paul, MN 55108 Email: ***@***.*** Phone: 773-458-3090 *Any work utilizing UIC equipment or staff support must acknowledge the University Imaging Centers in publications and presentations.*

On Mon, Sep 15, 2025 at 9:27 AM Cédric MESSAOUDI ***@***.***> wrote: *cmessaoudi* left a comment (MultimodalImagingCenter/MiC#5) <#5 (comment)> Which segmentation or detection algorithm are you currently evaluating? For example: Segmentation-focused: Cellpose, StarDist, U-Net (or variants) Detection-focused: R-CNN, YOLO, DTER, SAM To clarify: U-Net–based approaches (including Cellpose and StarDist) typically do not output per-object confidence scores, while object detection algorithms—whether standalone or combined with segmentation—usually do. You’re welcome to message me directly at my email address, which is publicly available online. — Reply to this email directly, view it on GitHub <#5 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AMB72WHYTSNZGB6ZQM4I6L33S3EANAVCNFSM6AAAAACDL7PJTKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTEOJSGQ2TSMJUGA> . You are receiving this because you authored the thread.Message ID: ***@***.***>