Understanding Seed based Confidence Connected SimpleITK Filter

Hi, I am using seed based ConfidenceConnected Filter to find the target of my interest in 3D image. The shape of target is different in each slice in the 3D image. I get much better results than Thresholding methods I tried before using seed based confidence connected approach. Couple of questions, I observed that confidence connected filter also detects some regions that are not completely connected and contiguous to the target of my interest. Is there any way I can filter non contiguous regions outside my main target of interest? The seeds I am selecting are well within the regions of my interest. Thanks Jiten.

Hello Jiten,

The connectivity is in 3D. So if you are just looking at a 2D slice you may not see how they are connected. I can recommend 3D Slicer for visualization. You can even use SimpleITK in the SimpleFilters module.

The results of the confidence connected filter are connected.You can use the connected component filter to give each connected region a different label id and sort by size.

Hope that helps.


Relabel filter is used for sorting and pruning connected components.


Thanks @blowekamp and @dzenanz I will try both approaches and share my feedback with results. Thanks, Jiten

Thanks @blowekamp and @dzenanz for your inputs. Since my target of interest takes different shape in each slice Post 3D segmentation step I apply relabelling on each slice by which I get more labels and can further filter regions by size. Another benefit I saw is by this I get more better bounding boxes at each slice level rather than bounding box in 3D space as shape my of target varies in each slice. Really serves my purpose.Thanks Jiten


How are you selecting the following?

The seeds I am selecting are well within the regions of my interest
Do you have an automatic method to create your seeds?

Hi @keesh, based on physical property of scan and location of target in the 3D scan I find the centre slice and a single seed point of reference that will definitely belong to target object. Once I get this I select random seeds within a small cubic region. Then I give all these seeds to confidence connected algo which nicely detects the pixels belonging to this target even when target takes different shapes in different slices. Let me know if it helps. Thanks, Jiten