Manual adjustment of segmentation output

[Meddebma]

Thank you very much for this repository.
If there are some failed segmentation, is it possible to adjust it manually? as you can see in this example the psoas muscles are not fully segmented at the level of L5

[louisblankemeier]

Hi @Meddebma,

Thanks so much for the feedback! The current muscle and adipose tissue model was trained only at L3. Because of this, performance degrades as you move away from L3. We are updating this model to a model trained at T12 - L5. Hopefully, this model should be up within the next few weeks.

Best regards, Louis

[louisblankemeier]

Unfortunately, there's no way to correct the segmentations manually at this point. Others have also requested this so we can start to think about ways to do this.

[Meddebma]

Hi @louisblankemeier,

thanks for the explanation, Now I understand it and I will be very happy to get the update!!! Why did you take 5 different slices? Most recent publications took only one slice per CT scan at L3 level. Did you get better results?

Thanks

[louisblankemeier]

Hi @Meddebma ,

We included the 5 slices as we hypothesized that analyzing these measures at each of these levels will be more powerful for downstream tasks than just using L3 measures and this has been shown in some recent literature. But, we are also hoping that the community can help validate this! So, would be excited to see what you find if you look into this.

Thanks a lot!

-Louis

[Meddebma]

Hi @louisblankemeier,

Sure, please let me know once you have more accurate results, I have ready clinically annotated high quality data that would help me validate your method! It would be also amazing if you could make it for MRI body composition,

Best, Aymen

[louisblankemeier]

Sounds good to me. Let's leave the issue open and I will update once we've updated the model.

Would be interested in adding MRI as well. If you have a MRI model that you would like to contribute to the pipeline, we could discuss further!

[Meddebma]

I do have good MRI data with clinical annotation but still not segmented, we can definitely discuss it further!

[louisblankemeier]

Btw, @ad12 suggested that we output dicoms that would enable manual correction in a software like ITK-Snap or Horos. Will do this so you can correct the outputs manually.

[louisblankemeier]

Hi @Meddebma,

Very sorry for the extremely long delay. Things were held up for several reasons, but we have finally integrated a new muscle + adipose tissue model into the spine_muscle_adipose_tissue pipeline. This model was trained on annotations from T12 - L5, as opposed to just L3.

Best regards, Louis

[Meddebma]

Hi @louisblankemeier, thank you very much for the update, my dataset is almost ready and I think of training the model this month. I'll be happy to share my results with you once I did it!

Best, Aymen