Skip to content

LLSM generated 3D images: Compatibility with cpsam_v2 training data requirement  #1491

Open

Description

@s-udisha

Before you fill out this form:

Did you review the FAQ?

Did you look through previous (open AND closed) issues posted on GH?

Now fill this form out completely

Describe the bug

I'm fine-tuning cpsam_v2 on 3D simulated bacterial cell data generated using lattice light sheet microscopy (LLSM) acquisition principles. Two inherent LLSM characteristics raise questions about training data preparation:

  1. Unequal XY and Z-dimension: LLSM generates images with different dimension in XY and Z, usually Z-axis is smaller than X and Y (e.g., 256 × 256 × 64). The transformer expects 256 × 256 input for each slice. Is this a problem if YZ and ZX slices are not 256 × 256?

  2. De-skewing: We deskew the pixels of the raw LLSM images to make z-acquisition look like that of a confocal (top-down slicing as opposed to the angular slicing from the LLSM). Deskewing therefore results in the parallelogram shape shown below. This is due to the padding with zeros during deskewing. I want to know if these sharp boundaries in YZ and ZX slices would be a problem for the fine tuning?

Image Image

Activity

  1. changed the title [-][BUG] make_train for 3D data slice generation issue[/-] [+]LLSM generated 3D images: Compatibility with cpsam_v2 training data requirement [/+] on Sep 14, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions