Deep Learning on X ray images

Part 1 (2018)
1

Hello,

I am interested in interpreting X ray images using deep learning. Particularly Chest X ray images. I am using Montgomery TB dataset and Shenzhen TB dataset for this experiment. I am trying to predict whether an X ray image is normal or has TB (nodule) indicaitons.

I tried fine-tuning the ResNet models from lesson 1. So far the performance is not that great. ResNet on untreated images (converting gray to 3 channels) gives me accuracy in range of 68-70%. I tried histogram equalisation on X ray image and accuracy was in the range of 74-75%. With histogram equalisation as one of the channels of the image, the accuracy was in the range of 77-82%. Accuracies on validation set is comparable to training set accuracies.

  1. Are there any data preparation methods which are well suited for X ray images?
  2. Are there any different models which I should be trying?
  3. Are Imagenet models capable of extracting patterns from X ray images? If not, should I build my own model rather than using Imagenet models?
  4. The total set of images are 800 in number. Should I be doing data augmentation? What sort augmentation helps? (I can think of flipping the image on the horizontal axis and small random zooming in.)

Thanks in advance.

1 Like
2

The techniques for X ray and CT are quite similar, except X ray is a bit easier since only 2d. Therefore, I’d suggest checking out the kernels, solutions, etc to the 2017 Data Science Bowl, which used CT scans.

4 Likes
3

Thanks @jeremy

4

Check this out - https://stanfordmlgroup.github.io/projects/chexnet/

3 Likes
5

I wonder if pre-training on medical datasets like this would help for un-related problems. I just feel like ImageNet doesn’t perform super well for non-rgb images (understandably) and since this dataset doesn’t have any rgb images it may help if the classification problem is for non-rgb images even if its for classifying something like icebergs vs ships.

6

I am not sure if there is any learning that is transferable from xrays to some other non rgb data set such as icebergs, the high level visual features are very different.

7

did you get anything working on this? Looking at a few medical imaging datasets now.