Segmentation Testing

[liyihao76]

Hello and thank you for your excellent work. I downloaded and installed the code and tried to follow the process of the demo (default parameters) for bone segmentation inference, but the result is not very satisfactory.
I tested our model using the datasets : SK10(https://ski10.grand-challenge.org/) and SPIDER (https://spider.grand-challenge.org/data/).
I tested 2D slice auto-segmentation and 3D full volume auto-segmentation respectively, but it doesn't seem to work very well. Especially for 3D segmentation, the inference results do not see a continuous structure.
Do I need any other configurations to achieve the desired segmentation results? For example, data preprocessing, inference parameter settings (lower and upper percentile), better checkpoints, manual prompts?

[Guhanxue]

Hi, do you mind providing me your MRI sequence information, and sending me an example for each exam if possible that i could do a troubleshooting with you?

[Guhanxue]

BTW, just through a quick look. I guess the issue might be your data was preprocessed volumes which is not under a meaningful clinical orientation. (such as the x,y,z axis are transposed during dataset sharing.)
For example, A real sagittal view spine usually in:
, where the spine is vertical in the red window in 3D slicer. I am not sure how much it influence the results, but maybe there is one reason.

[dyollb]

I am getting similar (or worse) 3D results for T1 abdominal and head scans and T2 head scans. Is there any pre-processing that I should do to make it work better? e.g.

• intensity normalization?
• resampling (expected resolution in mm)?
• image data orientation (e.g. RAS)?
• smoothing?

I tested with

• the amos22 abdominal (ribs+vertebrae) MRI (subject 0579) and
• IXI head MRI data (subject 025), and both T1 and T2 were acquired using a 1.5T scanner.
  I did n4 bias correction but no other pre-processing.

[Nilser3]

Hello!

I have similar results