So following William's work, I re trained en nnU-Net after correcting 20 segmentations over the data-multi-subject dataset William used before. The idea was to improve the segmentation, dealing with RL axis edges undersegmented and with concavities due to compressions oversegmented.
Here is the dataset, the four last are used for testing.
Dataset001_nnUNet_first_data_with_comp.json
Here the process during the training, I could have used a bit more epochs since it was quite fast (12 hours moreless), but it already seems to be converging and almost reaching the asymptot.
Dice score on tested data gave a result of .97, Which is good if we now consider that the data is well segmented on ths new dataset.
I'll have to make a QC report on the whole data-multi-subject dataset. But yet here is a good example of the progress on a zero shot segmentation for both models:
Now I also tested the model on dcm-zurich, I think I'll have to diversify the data with the various dcm datasets on Gitea. Some of RMIs from dcm_zurich are pretty well segmented but most are'nt. The data is also quite different, but since nnU-Net architetcure adapts pretty well to diversity in biomedical imagery, I think it would be interesting to have a diverse dataset.
So following William's work, I re trained en nnU-Net after correcting 20 segmentations over the data-multi-subject dataset William used before. The idea was to improve the segmentation, dealing with RL axis edges undersegmented and with concavities due to compressions oversegmented. Here is the dataset, the four last are used for testing. Dataset001_nnUNet_first_data_with_comp.json
Here the process during the training, I could have used a bit more epochs since it was quite fast (12 hours moreless), but it already seems to be converging and almost reaching the asymptot.
Dice score on tested data gave a result of .97, Which is good if we now consider that the data is well segmented on ths new dataset.
I'll have to make a QC report on the whole data-multi-subject dataset. But yet here is a good example of the progress on a zero shot segmentation for both models:
Now I also tested the model on dcm-zurich, I think I'll have to diversify the data with the various dcm datasets on Gitea. Some of RMIs from dcm_zurich are pretty well segmented but most are'nt. The data is also quite different, but since nnU-Net architetcure adapts pretty well to diversity in biomedical imagery, I think it would be interesting to have a diverse dataset.