Update:
this draft PR is WIP, and so far I've implemented the class of the metric.
The implementation refers to:
https://github.com/TissueImageAnalytics/CoNIC/blob/main/misc/utils.py
and fixes some of its bugs, such as calculate when the predictions or labels do not have background (instance id = 0).

I discussed with @KumoLiu , and according to my understand of the original paper, the predictions should only have 1 channel, since for each pixel, it should only belong to one instance. Therefore, the ground truth of an image with shape [H, W] must have the shape [H, W], thus to be consistent with other metrics in MONAI, I limit the input into shape [B1HW], am I understand right? Hi @drbeh , may need your help to confirm it, thanks!

/black

Update: this draft PR is WIP, and so far I've implemented the class of the metric. The implementation refers to: https://github.com/TissueImageAnalytics/CoNIC/blob/main/misc/utils.py and fixes some of its bugs, such as calculate when the predictions or labels do not have background (instance id = 0).

I discussed with @KumoLiu , and according to my understand of the original paper, the predictions should only have 1 channel, since for each pixel, it should only belong to one instance. Therefore, the ground truth of an image with shape [H, W] must have the shape [H, W], thus to be consistent with other metrics in MONAI, I limit the input into shape [B1HW], am I understand right? Hi @drbeh , may need your help to confirm it, thanks!

Hi @yiheng-wang-nv, thanks for implementing PQ metrics.

You are right, in the paper, the PQ metrics is calculated per class for the input is one dimensional for each PQ calculation; however, these PQs are averaged afterwards to create the metrics that we can denote as mPQ, and this is the actual metrics that we need.

Unlike the rest of MONAI metrics that are designed for semantic segmentation, this metrics is used for instance segmentation and they can have multiple channels. Having said that, it still depends how the data is being represented, for instance look at the ground truth of this dataset, where one channel is for instances and one channel for classes: https://conic-challenge.grand-challenge.org/Data/

I believe we can have two classes, one for binary PQ metrics where it assumes only one-channel inputs, and another for multi-class PQ where it accepts two channels (one for instance and one for classes) and you call the binary PQ inside it for each class.
OR, maybe it is better to have one class but accepting two one-channel tensors (instance map and class map) so for the binary one the second tensor is not needed.

Please let me know if you have any question.

Hi @yiheng-wang-nv, please let me know when this PR is ready for review. Thanks

@yiheng-wang-nv, it would be great if this can output SQ (segmentation quality) and RQ (recognition quality) in addition to PQ which is the combination of them. Thanks

Hi @drbeh , I modified the metric implementation to align with CoNiC challenge's version, and if calling compute_panoptic_quality directly, users can also do binary PQ calculation. As for the MPQ, we can use PanopticQualityMetric class and with the default setting: reduction = "mean_batch", the metric will return three (1, N) tensors which stand for PQ, SQ and PR, where N equals to number of classes. The calculation is the same as in CoNiC's function.

Could you please help to re-review it? Thanks, and if the format meets the requirements, I'll add unit tests asap.

/black

/build

/build

/black

/build

/build

/build