Fine-tuned worse than TotalSegmentator

[nairouzshehata]

I wanted to fine-tune TotalSegmentator's model for aortic segmentation from MR because the aortic root is not very accurate and I have 200 samples that are segmented which I wanted to use for fine-tuning.

The steps I followed (my dataset id is 4):

1. nnUNetv2_plan_and_preprocess -d 4

2. I download the weights into the pre-processed folder and renamed plans.json to nnUNetPlans.json

3. Then run this notebook that moves the plans from dataset id 730 to dataset id 4 (mine - MR aorta)

4. Then run this notebook for fine tuning

5. Finally for inference, I run this

Sadly, that's a fine-tuned sample

The white mask is the non-fine-tuned totalsegmentator prediction, which is much better!

• I used the 1000epochs trainer btw Thanks!