Deep-Learning-with-PyTorch

Bridging CT segmentation and nodule candidate classification
413
~2^25 voxels ~2^10
candidates
~2^25 ~2^20
1. Segmentation
~2^10 ~2^5
3. nodule
claSsification
4. malignant
claSsification
2. Grouping
~2^5 ~2^1
Figure 14.4
each step
The steps of our end-to-end detection project, and the rough order of magnitude of data removed at
Let’s go through the steps in figure 14.4 in more detail:
1 Segmentation—Segmentation starts with the entire CT: hundreds of slices, or
about 33 million (2 25 ) voxels (give or take quite a lot). About 2 20 voxels are
flagged as being of interest; this is orders of magnitude smaller than the total
input, which means we’re throwing out 97% of the voxels (that’s the 2 25 on the
left leading to the X).
2 Grouping. While grouping doesn’t remove anything explicitly, it does reduce the
number of items we’re considering, since we consolidate voxels into nodule
candidates. The grouping produces about 1,000 candidates (2 10 ) from 1 million
voxels. A nodule of 16 × 16 × 2 voxels would have a total of 2 10 voxels. 2
3 Nodule classification. This process throws away the majority of the remaining ~2 10
items. From our thousands of nodule candidates, we’re left with tens of nodules:
about 2 5 .
4 Malignant classification. Finally, the malignancy classifier takes tens of nodules
(2 5 ) and finds the one or two (2 1 ) that are cancer.
2
The size of any given nodule is highly variable, obviously.