ChestNet -- An Amiga Neural Network Program for Diagnosis of Chest Diseases.

Introduction

• According to this model, the brain takes data (the INPUT layer), processes it in some manner (the HIDDEN layer), and then provides an answer (the OUTPUT layer).

• Before it can do this, it must, like a child, be TRAINED with known data (INPUT/OUTPUT pairs).

• After much repetition and many examples, the brain (HIDDEN LAYER or NETWORK) is trained -- at least to some level.

• The more training, the fewer the number of errors.

The beauty of a neural network is that once properly trained, it can provide answers even with inputs it has never seen. It is like our ability to see the familiar in new situations because of our life experience. Computers, so the thinking goes, may do it faster and better, especially when the number of variables is large. Computers also take fewer coffee breaks! Such is the nature of artificial intelligence.

ChestNet -- Amiga-based Neural Network Chest Diagnosis

• I used Helm (Eagle Tree Software) to create an interface to NeuroPro 2.0 (MegageM) and linked the two using AREXX.

• I trained NeuroPro to recognize several chest diseases, based on the clinical symptoms of the patient and the radiographic appearance of the chest. Because the neural matrix in NeuroPro is 256x256, I was limited to 32 ascii variables to distinguish among diseases.

• In my application, the user describes the patient and the chest x-ray by

  selecting buttons.

• The user then can then get ChestNet's opinion as to what ails the patient.

• It does this by running Neuro Pro 2.0.

• Other options include:
  • Finding out which diseases ChestNet knows about.

  • Learning characteristic findings in each disease.

  • Selecting radiographs and CT scans for correlation.
    • You could, for example, see what a CT scan of what
      pneumocystis pneumonia would look like.

• There is even a link to another application which demonstrates an algorithmic approach to chest diseases.

Possible Questions

Have you published your ChestNet application?

I have demonstrated all of my teaching applications at the 1995 Annual Meeting of the Society of Nuclear Medicine held in Minneapolis, MN. I have also submitted a manuscript to Amazing Amiga but it is up to them if and when to publish it.

Can this application be adapted to run on the Internet?

I think the answer is "yes."

HTML programming supports radio buttons, check boxes, menus and the like. This would replace Helm as the front end.

I would then need a neural network engine that can run on a Unix shell. Either Dan Wolf would need to re-compile his NeuroPro 2.0 program or I would need to find a replacement.

Finally, I would need to write a "cgi" script program that could transmit selections from the user and retrieve answers from the network.

Why would anyone one want to do this?

Suppose there a problem or a calculation that required so much computer resources that only a few computers in the world could handle it?

Suppose certain computers became "experts" in certain kinds of problems. You might, for example, want to ask only the "smartest" or most experienced computer its opinion on how to treat your patient.

Why hasn't this been done before?

Maybe somewhere it is being done. Besides, I want "The Amiga Radiologist Home Page" to be on the cutting edge!

revised -- December, 2002

List of Projects -->