20.03.2021 Views

Deep-Learning-with-PyTorch

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

406 CHAPTER 14 End-to-end nodule analysis, and where to go next

1. Nodule Candidate Generation

1a. Segmentation

n 1b. Grouping 1c. sample tuples

((...,(..., IRC),

...

(..., IRC))

2. Nodule and malignancy claSsification

2a. Nodule claSsification 2b. ROC/AUC Metrics 2c. fine-tuning malignancy model

3. End-to-end detection

3a. (..., IRC) 3b. Is nodule? 3c. is Malignant?

Figure 14.2

A detailed look at the work remaining for our end-to-end project

b Grouping—We will group the voxels into nodule candidates by applying a threshold

to the predictions, and then grouping connected regions of flagged voxels.

c Constructing sample tuples—Each identified nodule candidate will be used to

construct a sample tuple for classification. In particular, we need to produce

the coordinates (index, row, column) of that nodule’s center.

Once this is achieved, we will have an application that takes a raw CT scan from a

patient and produces a list of detected nodule candidates. Producing such a list is the

task in the LUNA challenge. If this project were to be used clinically (and we

reemphasize that our project should not be!), this nodule list would be suitable for

closer inspection by a doctor.

2 Classify nodules and malignancy. We’ll take the nodule candidates we just produced

and pass them to the candidate classification step we implemented in

chapter 12, and then perform malignancy detection on the candidates flagged

as nodules:

a Nodule classification—Each nodule candidate from segmentation and grouping

will be classified as either nodule or non-nodule. Doing so will allow us to

screen out the many normal anatomical structures flagged by our segmentation

process.

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!