suggestions to overcome over segmentation

[Aseem139]

Hi All,

I have been using Biapy for instance segmenation. I an really happy with the segmentations but I think that masks could be a little smaller. Below is what the model currently predicts

I am attaching the yaml file as well. Any help would be really appreciated. Thanks in advance.

2d_instseg_train.yaml

[danifranco]

Hi,

If you are using the last release (3.6.7) you can modify the PROBLEM.INSTANCE_SEG section as follows:

PROBLEM:
    INSTANCE_SEG:
        DATA_CHANNELS: ['F', 'C']
        WATERSHED:
          SEED_CHANNELS: ["F", 'C']
          SEED_CHANNELS_THRESH: ["auto", "auto"]
          TOPOGRAPHIC_SURFACE_CHANNEL: "F"
          GROWTH_MASK_CHANNELS: ["F"]
          GROWTH_MASK_CHANNELS_THRESH: ["auto"]

This is the new way of creating the instance segmentations (we still need to update the doc/templates to reflect these changes). However, the "old" configuration you are using is being translated into this new setting automatically, as we always ensure backwards compatibility between configurations. As you can see in this new configuration, the watershed process is more explicit and customizable. Currently, for creating the seeds, F (the old B channel) and C channels are being used (SEED_CHANNELS). As these channels are in 0-1 range, we need to threshold them to identify the centers of the instances. Ideally, we will avoid touching instances by removing the countour pixels to the foreground ones. The probabilities to threshold them are being automatically calculated using otsu ("auto" in SEED_CHANNELS_THRESH for each of the channels). Once that the seeds are created they will be grown by the watershed constrained by a growth mask. This growth mask is created by thresholding F channel (GROWTH_MASK_CHANNELS) automatically too, see the "auto" in GROWTH_MASK_CHANNELS_THRESHvariable. So, to control how large are the final instances, or in other words, to control how much the seeds will grow, you would need to control that last GROWTH_MASK_CHANNELS_THRESH. You can set it with a float instead of "auto", e. g. 0.7` to constraint a bit the created growth mask. A lower value will imply a larger mask, which will imply creating larger instances.

I have modified the yaml for you here:
2d_instseg_train.yaml

The above explains what to modify to alter the watershed process. However, apart from that there is also a instance refinement phase right after the instances are created by the watershed and before any other post processing. Here is the link to those variables:

BiaPy/biapy/config/config.py

Line 1646 in 3fc05ae

_C.TEST.POST_PROCESSING.INSTANCE_REFINEMENT = CN()

. There, as you may see, you can apply morphological operations to the instances such as erosion, dilation, fill holes etc. For example, to erode the instances you would require something like this:

TEST:
    POST_PROCESSING:
        INSTANCE_REFINEMENT:
            ENABLE: True
            OPERATIONS: ['erosion']
            VALUES: [3]

[Aseem139]

Thanks a lot. The works really well.

[danifranco]

Hi!
Just to mention that we have updated the instance segmentation documentation to the current approach in case you want to check it.

[Aseem139]

Hi Daniel,

Thanks a lot for updating the documentation. Just finished reading it and it's amazing. Do you have any suggestions for the channels to choose when segmenting mitochondria, given the diversity in shape and size of different instances?

Also i have a quick question. I am trying to run predictions using I model I have trained. I would like to use the pretrained weights and make predictions on a folder of images. Here is the yaml I am using currently but it does not work. I am attaching the yaml and the error.

2d_instseg_pred.yaml

I was expecting this to run prediction on the images present in folder with the location ....TEST.PATH

[danifranco]

Hi,

F + C should work for mitochondria as well, that combination is quite generic.

Regarding the issue will try to reproduce and fix it. Are you using the code of the last release or the last commit?

[danifranco]

I have checked and there was a bug defining PATHS.CHECKPOINT_FILE as you mentioned. I have fixed it for the next release. here 236d68d .
In you specific case, you can avoid setting that variable and putting the checkpoint file a the place where BiaPy will look. Looking at your screenshot, if your result_dir is /home/akas0018/cm10/biapy/instance_segmentation/stardist_2d/inference_results placing the checkpoint at /home/akas0018/cm10/biapy/instance_segmentation/stardist_2d/inference_results/2d_instseg_attention_unet_bcd/checkpoints/2d_instseg_attention_unet_bcd_1-checkpoint-best.pth should do the trick!

[danifranco]

I'm closing this as it is solved but feel free to reopen it or create a new issue if you face any error.