Status update for Cytoseg

[Rick Giuly]

Automatic Segmentation

Using two major stages:

(1) A denoising stage that acts on local patches - it has the effect of
cleaning up the image and enhancing contrast.

(2) Contour detection in 2D and linking contours to form 3D objects.

(All groups working on automatic segmentation of neuropil use some form
of (1), (2), or both, but the details differ for each implementation.)

Denoising

Some use a neural network classifier for denoising. Currently, Cytoseg
is using a random forest classifier, which is different in
implementation but accomplishes a similar task.

A new feature is the ability to enhance more than just one structure,
selectively. For example, you get one output that represents membranes
and another output that represents mitochondria.

There are some examples of this posted here:
http://cytoseg.blogspot.com/

New functions have been written to calculate accuracy, also part of the
cytoseg project code.

Here are some numbers for current voxel detection accuracy:
mitochondria 91% true positives and 15% false positives
membranes 90% true positives and 16% false positives

Contour Detection and Linking

Once the images are denoised, it's possible to extract reasonably clean
contours with simple algorithms. Those contours then have to be linked
in 3D. Future work will be on methods to reliably detect that contours
in two adjacent planes should be grouped together (because they came
from the same 3D object). The simplest way is checking the area of overlap.

(There's also more to explore in the way of using properties of the
detected contours to determine the type of objects more accurately.)

Manual Segmentation Efforts to Generate Training Data

People can process these SBFSEM datasets much more quickly with Amira
than with Imod (or equivalent applications). Amira provides a
thresholding tool that gives a crude segmentation to start with and then
he can clean it up with the rest of Amira's tools, which are generally
well-designed. The training sets can be stored as simple tiff stacks.
Amira uses raster maps to represent segmentation, so you end up with an
image stack where the number associated with each voxel represents a
particular object index, rather than a grayscale value.

It is possible to combine manual and automatic segmentation. For
example, we can feed the denoised output of the automatic segmentation
back into Amira and clean it up. There are some implementation details
that will have to be worked out before this actually happens.

Automatic Segmentation Speed

Automatic segmentation is extremely demanding in terms of CPU power. It
takes about 10 minutes to process a 175x175x2 image stack. To process a
1000x1000x400 stack would obviously take quite a bit of time, so we will
need parallel processing ability to handle this eventually.