davidchall / geodosic

Finding geometric patterns in dose distributions

Skills 🛠

• Languages: R, Python, SQL, C++, MATLAB
• Big Data: Spark, PrestoDB, Azure, AWS
• DevOps: Git, Docker, CI (GitHub Actions, Travis, Jenkins)

geodosic's Issues

Handle nested polygons when creating structure mask

Need to determine hierarchy of polygons in a z slice, and then assign the hierarchy level to each polygon (i.e. 0 for largest polygons, 1 for first level of nested polygons, etc).

• polygons with even hierarchy level contribute True to the mask
• polygons with odd hierarchy level contribute False to the mask

Store intermediate results efficiently

Currently serialize intermediate results using msgpack_numpy, but this produces large files.

Possible solutions would be:

• numpy.packbits()
• numpy.savez_compressed()
• hdf5

Use np.logical_xor in structure rasterization

I currently use a checkerboard algorithm which multiplies by -1 each time a voxel is included in a contour, followed by a subsequent boolean conversion. The resulting code is fairly complex.

I think this could be replaced by a simple call to np.logical_xor.

Structure rasterization fails when separated in z

When a structure contains distinct and separate superior and inferior substructures, the rasterization fails. See the comparison between the geodosic mask and the CERR contour below:

There is an inferior substructure not visible in this slice. This causes the bottom slice of the superior substructure to be extended further than it should. This looks like a regular tube at the bottom of the structure.

I think this is because the structure_mask function assumes there is a single structure when calculating z_min and z_max.

Serialize intermediate results

Messagepack provides a binary JSON-like data format. Would need to pip install msgpack-numpy. It should install via Cython and then provide Python bindings. It provides a similar interface to the pickle and json modules, if I use dumps() and loads().

import msgpack
import msgpack_numpy as m
m.patch()

Structure masks can be stored more efficiently:

b = np.packbits(a, axis=1)
c = np.unpackbits(b, axis=1)

This axis=1 should also retain the shape of array a.

Invalid rasterization when contours are spread out

Currently the rasterization function iterates through the z planes of the rasterization grid. If it cannot find a contour plane that is closest to the grid plane (i.e. both planes identify the other as their closest plane) then it doesn't rasterize the structure on this plane. This is how we find gaps in structures.

If the contouring grid has a larger delta z than the rasterization grid, then the above test will identify gaps in the structure, when in fact we should be interpolating the structure between these gaps.

This is unlikely to become a problem in practice because contouring is done on a CT or MRI, which tends to have a grid with the smallest delta z. Dose is always scored on a coarser grid.

Check grids are uniform before calculating distance transform

Would be better to have a function in geometry module that takes a grid as an input, instead of a DicomRtDose method.

Model scoring should support multiple metrics

Currently, model scoring will score the model based upon a single metric. We loop over metrics in an external loop in the IPython notebook, and this means that the predicted and planned DVHs are calculated each time.

It would be more efficient to enable model scoring to handle multiple metrics, so the DVHs are only computed once. Plotting would use PdfPages in this case.

Do I want to make each metric a function that takes a DVH as an input, or keep using DVH methods?