Challenge: Complete a tibia segmentation within one "Turkish march" (Mozart)
Testing out the #opensource medical image segmentation code I'm working on. Algorithm:
- Click a voxel
- Get voxel coordinates + intensity
- Get the nearest intensity contour line
Next one can edit and combine these contours with operations like cut/merge/smooth/... , and... if all else fails manually draw things.
For this video I set myself the challenge of segmenting a tibia from a crummy MRI within minutes. The tibia's outer contour is dark (hardly any MRI signal) while its core (marrow/spongy bone I believe) looks white. This used to take me 10 times longer but I've added a basic "prediction" of contours for the next slice. It can be improved still as you can see but when the data is well-behaved one could accept the prediction for most slices. Here you see I sometimes have to remove some "sticky-outy" bits (e.g. connective tissue bits at the bone that also look dark in MRI). Also, as I move up these slices, the data gets increasingly noisy (the leg moves away from the acquisition coil), and the dark cortical bone region gets thinner. So you can see me run into some trouble there and have to resort to more manual editing, and I'm relying (a bit too much at times) on special smoothing splines to iron things out.
I'll be adding this to #Comodo, which is the #JuliaLang project I'm working on, and hopefully in time for the Comodo workhshop at #cmbbe2025 CMBBE Symposium
https://github.com/COMODO-research/Comodo.jl