vault backup: 2025-01-11 22:08:25

HarryHeres · Jan 11, 2025 · 0ea85ac · 0ea85ac
1 parent 47c1600
commit 0ea85ac
Show file tree

Hide file tree

Showing 11 changed files with 89 additions and 41 deletions.
diff --git a/docs/Notes/Pasted image 20250110180214.png b/docs/Notes/Pasted image 20250110180214.png
diff --git a/docs/Notes/Pasted image 20250110180302.png b/docs/Notes/Pasted image 20250110180302.png
diff --git a/docs/Notes/Pasted image 20250110180350.png b/docs/Notes/Pasted image 20250110180350.png
diff --git a/docs/Notes/Pasted image 20250110180459.png b/docs/Notes/Pasted image 20250110180459.png
diff --git a/docs/Notes/Pasted image 20250110180620.png b/docs/Notes/Pasted image 20250110180620.png
diff --git a/docs/Notes/Pasted image 20250110180719.png b/docs/Notes/Pasted image 20250110180719.png
diff --git a/docs/Notes/Pasted image 20250111203551.png b/docs/Notes/Pasted image 20250111203551.png
diff --git a/docs/Notes/Pasted image 20250111210107.png b/docs/Notes/Pasted image 20250111210107.png
diff --git a/docs/Notes/Pasted image 20250111220818.png b/docs/Notes/Pasted image 20250111220818.png
diff --git a/docs/Notes/To Do and Meeting Summaries.md b/docs/Notes/To Do and Meeting Summaries.md
@@ -0,0 +1,89 @@
+### TO DO and Meeting Summary January '25
+#### Summary
+- Arthur successfully implemented SSM pipeline 
+- works for partial meshes, ignoring points that are far away (achieved with RANSAC Regressor)
+- distal reconstruction in examples looks good
+- Arthur noted slight offset proximal rotational alignment would cause more distance offset distally
+- Arthur tested different number of landmarks - in future can also tests effects landmark locations
+- tps transform faster with fewer landmarks
+	- note when changing # landmarks need to update number of PCs
+
+**question from Eva**> should holes be filled automatically or no? we don't want to add extra vertices in there since it might mess with algorithm right? maybe fill with faces *after* resampling?
+#### TO DO
+**Eva:** 
+- [ ] test SSM pipeline 
+- [ ] compare outputs to paper using same dataset and SSMs
+- [ ] bring Max onto project to do further landmarking to increase dataset?
+- [ ] update remeshing functions to work with newest Slicer, and add optional smoothing
+- [ ] testing RANSAC alignment criteria? how "strict" can we go? develop ideas starting point and then increments to increase by until alignment can be achieved
+
+**Jan:**
+- [ ] if time, RANSAC alignment optimization would be great
+
+
+### Eva summary Ransac distance threshold tests
+
+- default 2.7%, repeat 2x to check variation between runs
+- ![[Pasted image 20250110180214.png]]
+
+
+- here relative to target (magenta)
+- ![[Pasted image 20250110180302.png]]
+
+- Ransac = 1 (long axis rotation worse than 2.7 but rest is better aligned, same with distal alignment bc of rotation about other axis)
+
+![[Pasted image 20250110180350.png]]
+green is 2.7, yellow is 1
+![[Pasted image 20250111203551.png]]
+earlier test from RANSAC = 1 (better rotation alignment)
+
+![[Pasted image 20250110180620.png]]
+
+
+Ransac 0.5 in blue - again worse long axis rotation than higher ransac so overall still probably too high ransac value:) zellow was ransac ^1=
+
+![[Pasted image 20250111210107.png]]
+
+![[Pasted image 20250111220818.png]]
+### TO DO and Meeting Summary Dec '24
+Eva
+- [x] upsample meshes
+	- [x] fill holes
+	- [x] upsample
+	- [x] clean
+- [x] double check point numbers pre and post crop to make sure no issues 
+
+Jan
+- [x] add scaling factor in GUI
+	- in global_registration:
+		- Arthur used: 
+		-  `distance_threshold = voxel_size * 1.5 
+		- voxel_size = size / 55 (size is bounding box)`
+		-  so our default distance_threshold in GUI should be size*1.5/55 
+
+
+	- so to replicate Arthur's settings
+		- if we do 0.007 factor (0.7%) of bounding box size for distance_threshold for ICP
+		- then 0.027 (2.7%) of bounding box size for distance_threshold for Ransac
+		- can interconvert, I think it makes sense for user to put it goal for ICP since that is the final one, and calculate Ransac from that
+
+
+
+*Idea for setting max deformation allowed*
+
+- check after rigid alignment max distance between two corresponding points (on cropped models only)
+	- could use DeCa for this since the models are already aligned
+- then use that to inform lambda value for deformation 
+	- ie model allowed to deform enough to match proximal surfaces but distal surface can only deform by this amount
+
+
+
+*Other notes from Meeting 29/11/24*
+- we discussed that SSMs will not solve our problem (as Jan already stated in thesis defense)
+	- gives variation in length but would not necessarily help predict where within that feasible space the individual falls
+		- indeed, other SSM methods for distal humerus reconstruction have issues with getting length exact
+	- we would be happy to get similar errors to them
+	- but we can do so by limiting deformation as discussed above
+		- true length could only really be predicted if some part of proximal morphology can be used as predictor of distal
+- future idea find closest match in proximal morphology (after scaling and rigid alignment)
+	- check if finding closest match from "libary" of models used to create mean model gives better results than using mean model
diff --git a/docs/Notes/To_Do.md b/docs/Notes/To_Do.md