HTM for nodule detection in CT Sequences


I’m very much a HTM newbie and trying to read as many papers as I can, so go easy on me!

I’d like to experiment with HTM on CT scan sequences of lungs, to see if it can detect a nodule (tumor) within a CT scan. I’m aware you can achieve this with CNN.

A CT scan involves taking a series X-ray images (slices) from the top to the bottom of an organ.
For example (CT on lungs):

I was thinking:

  • a full CT scan (comprised of all slices - from the top of the lungs to the bottom) could represent a single pattern to be learned
  • make the CT scan artificially temporal by transforming it from series of images into video format
  • do any necessary preprocessing - binark mask, normalisation
  • encode the data into SDRs and feed to the temporal pooler
  • after training, introduce a CT scan with an anomaly (nodule)


  • could this work even if it is artifically temporal? will it not work because HTM predicts at every time step?
  • would there be too much variation between scans (scans from different people, not the same pair of lungs every time - could it still learn the pattern?)
  • how many training instances would it require?
  • would this require some serious resources to run, or will that depend on how much I can reduce the size of the input data?

Thank you very much in advance!


Hello @dee and welcome to the community :slight_smile:. Are you trying to find a way to use HTM that will perform better than state of the art machine learning techniques? If so, I suggest you find another problem. The CT scan problem you’ve defined is purely a spatial search problem, the type of problem Deep Networks work very well on. I don’t think you’ll find a better cancer detector with HTM that you will with todays DL systems.

To create an HTM solution for this, judging from the sample video you posted, the input space would be very large, too large to be very efficient.

The artificial temporality of the data could prevent you from identifying where in the organ an anomaly was found. There are tactics for trying something like this, but I doubt you’d get better results than state of the art deep networks.

Sorry to be naysayer, but this just isn’t want HTM is good for. :man_shrugging:


Hey @dee, welcome! Regarding your application I agree that this is the issue:

While CNNs are equipped to find spatial features, HTM is built for temporal features. So HTM is better fit for other tasks like detecting irregular heartbeats, since the anomalous activity is expressed over time in the form of unusual rhythms.


Hi rhyolight,

Thank you for replying so quickly! Your answer makes a lot of sense. I’ll take your advice and look into other things. If you have any ideas or know where I could look, please let me know.

1 Like

Well I took Andrew Ng’s Coursera course, and it was worth the money. I also hear the courses are very good, but I haven’t tried them.

Hi sheiser1,

Thanks for responding! I agree, it makes more sense to me now. Yes I’ve read some papers on it’s application to ECG signals, very cool! I might look into using HTM on EEG data

1 Like

Sorry I mean’t HTM ideas!

1 Like