heading

Segmentation Methods

ND320 C3 L3 06 Segmentation Methods

Quiz: Can you identify a use case for segmentation in this clinical task?

QUESTION:

Selecting the right method

A certain part of the population has risk factors that make them susceptible to early lung cancer - these could be things like smoking, routine exposure to certain substances or family history. A combination of these risk factors make people good candidates for routine lung cancer screening since early detection can lead to very positive outcomes. Routine lung cancer screening is done by taking a low-dose CT image and then looking for dense areas in the lungs, or lung nodules. Quite often, if lung nodules are found, they need to be monitored to see how they grow as the presence of nodules per se does not necessarily mean that intervention is needed. Most important question a radiologist would need to answer - which nodules have increased in size since the last time a scan was taken?

Would a classification or segmentation algorithm be a good tool to assist the radiologist? Why?

Write down your thoughts below.

ANSWER:

Thanks for your response. The segmentation algorithm would make the radiologists’ job much easier since measuring volume needs accurate delineation of the extent of the nodules. However, in a comprehensive AI system, a classification or object detection algorithm may also assist in the screening process, providing a second read of the image and flagging those where nodules are initially found.

Summary & Exercise Instructions

A U-Net architecture has been very successful in analyzing 3D medical images and has spawned multiple offshoots. You will get a chance to get more familiar with it in the exercise that follows, but if you would like to understand the principles better, I recommend that you check out the webpage on U-net created by one of the authors of the original paper, Olaf Ronneberger: https://lmb.informatik.uni-freiburg.de/people/ronneber/u-net/index.html. You will find the link to the original paper and a few materials explaining how and why this architecture works.

Now, let’s move on to the exercise where you will have a chance to train your own segmentation network.

Next Concept