search

sha168 / ADNet

Code for the paper "Anomaly Detection-Inspired Few-Shot Medical Image Segmentation Through Self-Supervision With Supervoxels".
MIT License
27 stars 10 forks source link

A problem of Supervoxel generation #3

Open lijing-coder opened 1 year ago

lijing-coder commented 1 year ago

Supervoxels were generated for 3D data in AD-Net, but when I displayed the images in the generated Supervoxels, the quality of the generated Supervoxels was not good in terms of pure visual perception. I also displayed the superpix of ouyang under the same image, which felt good. Specific visible images.

image image a982f95ec063a13b31d8c9c7b07c9dac

This picture is the 33rd image of image_37.nii.gz, the top two images are taken from the supervoxel generated by AD-Net, and the following images are the superpix generated by ouyang.

This picture is the ground truth. image

lijing-coder commented 1 year ago

Have you ever compared using the superpix generated by ouyang directly for AD-Net training?

lijing-coder commented 1 year ago

Thank you for your excellent work and look forward to your answer!

shenqq377 commented 11 months ago

Have you ever compared using the superpix generated by ouyang directly for AD-Net training?

Yeah, We have found the similar problem in some hard cases when you process 3D data. We didn't compare these two approaches. For a certain organ, we get different superpixel class ids on different slices by using superpixel segmentation slice by slice, thus we are unable to select adjacent 3 slices based on a specific superpixel classes id. It is a good idea to do the comparison.