Identifying Tuberculosis with Artificial Intelligence (AI)

We’ve started experimenting with AI, and it’s very interesting to see the results. The use case for this experiment is identifying TB based off of 100 images.

Google AutoML Vision

Google’s AI platform is a decent but expensive platform. It costs nearly $200 to use their AutoML Vision Beta, but it’s worth it if you don’t want to deploy training Python algorithms on local or in the cloud. Essentially, you will use Google algorithms to train images, and it’s getting to be user friendly. We provided some feedback to them, but they probably won’t add to their next release…

The other catch to the $200 startup fee is the training time. You get 1 hour free of training time, so if you have a lot of labeled images, you will be charged after an hour. Once you click train (100 positive and 100 negative images is recommended), the clock starts ticking. We weren’t charged since we labeled 100 total positive and negative TB images.

The interesting thing is, the accuracy was still pretty high. From the screenshot above, 87.5% precision was recorded. After we discussed the potential of AI in medicine, a doctor friend said, “…AI is taking my job away,” which should be an interesting topic for future elections…

This is an interesting use case which can be applied to anything really, and imagine if we had more than 100 images trained. Theoretically, the accuracy should be even higher.

If you have any questions or comments, please email dev [a] spacegoodies [dot] com