ajensen1234 / jtml Goto Github PK

I want to tease out the optimizer so that the underlying functionality is separate from the main body. Remove all the copy and pasted code.

Want to be able to move the models at a slower keyboard speed than 1deg or 1mm.

When changing the frame you are looking at on Linux, the update rate is extremely slow (1 change/sec). This can make it difficult to see the movement of the implants throughout the movement and to quickly check images.

When optimizing, I like to look at the dilated edges. However, once you have gone to the dilated edges, clicking on the “original” image doesn't return to the original image.

We want to be able to use the same branch, along with some CMake flags, so that we can have singular compilation across platforms. This will drastically improve the ability to push and publish for different operating systems, and make testing a bajillion times easier across operating systesms.

Instead of using the Hamming distance as a cost function (which has no notion of "wrongness" if there is no overlap, we want to implement a distance map as a cost function.

We think that this has the same strengths as the Hamming distance, but also encourages a bit of convexity in the search space, which should allow multiple iterations of the optimizer to find the true location. We shall see...

Probably going to be a perpetual issue, but I'm creating this to help track changes. Branch will be called cleanup-main.

I think this might have to do with VTK control.

• Make dummy actor
• Disable VTK control

We want to add a small box that shows the current 3D views of the implants in the coronal (frontal) plane. This makes it easier to determine if the implant has registered correctly. Makes it easier to view the varus/valgus

Optimizer places things in the correct location, but the front end VTK viewer is moving things to the wrong place.

I'm guessing that something is wrong with the way the window center is being placed.

The algorithm works, but not with the new additions to the viewer and different locations for each of the implants to live. Need to fix that so we can perform different calculations

Loading in a second model after an initial model has been loaded breaks the system. The second model is not placed correctly.

It is the correct color however

See title.

When selecting the model you wish to view in the model selection area, the currently selected and not currently selected items are not showing correctly. I have UF_ORANGE and UF_BLUE hex codes saved, just need to get the gui to display that correctly.

Need to edit some of the back-end Qt files that we used for creating the front end ui to move the optimize backward button into the same section as the rest of the optimizers.

There are some calibration hacks in the background that assume the input image is always going to be 1024x1024. Need to fix those to read in the image size once loaded and make adjustments based on that input.

Might need to do a little bit of a rewrite of the calibration backend.

Might be easier/more reasonable to do this once #26 has been completed.

Qt5 was originally released in 2012. Even though 5.15 is LTS, I think it would be better to move over to newer versions. I don't think we are using any modules that will drastically break, other than MAYBE the OpenGL renditions.

When using multimodel selection on Ubuntu, It merely acts as a single selection

Any reference of loaded models or model actor list will break things. Fix this up so we can accept the correct datatype

With shoulder implants, we are running into the issue where the salient regions of the implant don't seem to have enough “oomph” to get the implant optimizing to the right place.

With this, we want to find the regions of high curvature in the CNN output, then implement the Hausdorff distance transform for each of those points. Because these points are salient to human supervised version of aligning implants, we believe that this will be a good heuristic for model-image registration.

The number of iterations per second displayed in Linux is broken, despite the optimizer running at the correct speed. We can tell because the total number of iterations is correct.

We want to be able to build on both Linux and Windows machines. Probably using GitHub actions for now.

• Get BA plot from UWestern
• Create calibration constant for VV adjustment based on specific input measurement
• Enfroce VV correction based on loaded calibration constraint

I'm pretty sure it's possible for each separate widget on the mainscreen to have a separate header file. This would majorly declutter the mainscreen.cpp file for ease of use.

Want to train a decision tree neural network to help pick the correct pose of a tibial implant.

Need to train something on PyTorch (or something else that can get loaded into C++, maybe ONNX?).

Train on the kinematics from all the data that we currently have, inputting both the correct pose and the symmetric pose, along with the femoral pose. Train some decision tree to find the correct pose.

We should be able to parse through the directory of cost function DLLs instead of needing to load everything into a single file. I.e. make it easier to write your own cost function for people to share and use.

Program crashes when dr banks loads neural network. Might be an issue with deprecated pytorch function for loading NN models.

Doesn't crash on other computers - strange.

Want to add documentation for what each of the different optimizer directives are doing