How to tell ML to avoid previously labeled segmentations?

[ross-mitchell]

Hello:

I am segmenting several objects in the same scene (multiple organs in abdominal CT scans). There are situations where it would be very helpful to tell MONAI Label to avoid already labeled structures (besides background). That is, tell ML not to grow into voxel regions that already have a non-background label.

Is this easily doable?

[danieltudosiu]

Hi Ross, To do that you would need to integrate some sort of partial update of the segmentations. At the top of my mind come those papers that might be of interest: Lin, Z., Duan, Z.P., Zhang, Z., Guo, C.L. and Cheng, M.M., 2022. FocusCut: Diving Into a Focus View in Interactive Segmentation. In *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition* (pp. 2637-2646). Chen, X., Zhao, Z., Zhang, Y., Duan, M., Qi, D. and Zhao, H., 2022. FocalClick: Towards Practical Interactive Image Segmentation. In *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition* (pp. 1300-1309). Sofiiuk, K., Petrov, I., Barinova, O. and Konushin, A., 2020. f-brs: Rethinking backpropagating refinement for interactive segmentation. In *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition* (pp. 8623-8632). About the ease of implementation, it all depends on how familiar you are with MONAI and Ignite libraries, and to some degree with PyTorch Lightning due to similarities to MONAI Label. From my skimming of the papers, it would require some sort of data cropping based on preexisting segmentations. Sadly I am not that familiar with Datastore and how MONAI Label handles data processing to answer you in detail. Hope it helps, Dan

…

On Mon, 14 Nov 2022 at 21:25, Ross Mitchell ***@***.***> wrote: Hello: I am segmenting several objects in the same scene (multiple organs in abdominal CT scans). There are situations where it would be very helpful to tell MONAI Label to avoid already labeled structures (besides background). That is, tell ML not to grow into voxel regions that already have a non-background label. Is this easily doable? — Reply to this email directly, view it on GitHub <#1145>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACQEEVE4RJML7D4JRLJSQODWIKU25ANCNFSM6AAAAAASAICTWU> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[diazandr3s]

Hello:

I am segmenting several objects in the same scene (multiple organs in abdominal CT scans). There are situations where it would be very helpful to tell MONAI Label to avoid already labeled structures (besides background). That is, tell ML not to grow into voxel regions that already have a non-background label.

Is this easily doable?

Hi @ross-mitchell,

It is not easy to tell the MONAI Label model to not grow into custom regions. It'll always get the predictions for all the labels it was trained on. Unless you want to implement what @danieltudosiu recommends.
You could get all the predictions from the pretrained and then modify them in the viewer - all this is more of a viewer operation I'd say.

Hope this makes sense. Happy to discuss this further in a videocall if needed.