Automated Medical Image Analysis

[cogmandino]

---

Members do not see this ad.

I am starting a research that involves integrating Image Analysis programs with medical images such as X-Rays or CT scans.

I would just like to know what kind of diseases can be diagnosed/confirmed with medical imaging. What are the usual things medical images capture? Are there certain problems that are difficult to see in medical images but can be spotted?

For now, we are planning to use Image Analysis for detection of Community Acquired Pneumonia from Chest X-Rays. However, we would also like to know if there are other kinds of diseases that we can diagnose or other medical images we can use.

Thank you.

 

[colbgw02]

At present, CAD is mostly used to help detect breast cancer in mammography, and I think some people have been playing around with using it to detect nodules on chest radiographs. Without knowing anything about your research, I'm having a hard time figuring out how or why you'd go about detecting CAP. CAD isn't that good even for relatively straightforward things, much less for pulmonary opacities that have no distinct morphology and are extraordinarly nonspecific.

 

[jake2]

It's a long, long way off for almost all applications. It is used some in mammography and there are algorithms proposed for detected lung nodules and small pulmonary emboli. However, it is a notoriously intractable problem for almost all other applications.

For a baseline, consider that it's computationally not feasible for computer algorithms to read those fuzzy letters you enter when registering for a web site. By comparison, most radiologic findings take humans years to learn how to properly analyze.

As time goes to infinity, computers will not only be able to replace radiologists, but will become more intelligent than humans. That world probably ain't coming for a while though. 🙂

 

[durfen]

I have been in this field for quite a few years and I have to say that while computers can assist humans in analyzing images, the sheer variety of appearances of even the most basic finding will be a constant thorn in the usual programmer's backside.

However, there is one way. If the computer thinks like a human, endlessly fault-tolerant, flexible, and with a capability to see the world in gray, then computers can replace humans. They will have better error rates as well, since they will see more studies as they can be on call all the time. This is where neural networks and support vector machines come in. But then, you have to define the world for them. We undergo years and years of training to understand the world around us, so how do we programmatically define the world for an interpreter? Very, very difficult (but we only need to do it once). (I'm being very hand-wavy here without talking about tech details.....)

Remember though, if we forget how to interpret our own imaging, we can halt progress forever (therefore complete replacement will never happen).

And dude, eval for CAP? Most of the time its imaging+clinical, the patient is often hypovent, rotated, have lines and tubes running every which way. Even with humans, its always PNA versus atelectasis. In addition, people coming in with r/o CAP have a different problem very often. Consolidatoins/silhouette signs are the easy stuff, don't need image processing, the referring doctor can see it let alone a rad. Good luck, I would like to know how you want to approach it.

 

[johnfaust]

I have experience assisting a medical technologist in one laboratory as computer technician that time. I am interested to know how accurate it is when it comes to image analysis through computer.