Say we have microscope images obtained at various known magnifications. And I'd like the model to be able to tell that a big feature in a high magnification image is the same as a small feature in a low magnification image. In addition, after training using 50X and 100X magnification images, I'd like to model to also produce good predictions with a 70X magnification image. Intuitively, I'd think that feeding the model the concept of 'pixel size' is the straightforward method. Is there anyway to realise this? Or is it a readily available parameter in image related machine learning models? Or perhaps there is a better way to achieve the task?
I could easily be wrong, but I think the easiest way to go about this is to use the variation in number of features caused by magnification and scale accordingly. When you magnify/zoom in on an image, less of the image appears, right? So, you would have fewer features to extract. And since pixel size is held at 0.26mm, I don't think overriding with some new definition of a pixel would be very easy or pretty.
If I'm reading this right, I think your best bet would be to get fancy with scaling and do something like this: https://medium.com/neuronio/how-to-deal-with-image-resizing-in-deep-learning-e5177fad7d89
For dimensional reduction or images resizing, I suggest you follow https://stats.stackexchange.com/questions/192722/dimension-reduction-using-pca-in-matlab
If you are Python user, please look at https://www.datacamp.com/community/tutorials/principal-component-analysis-in-python
You can use convolutional layers to make magnification or resizing on images.
Hello Xiaobing,
There are many different ways to approach the problem of varying magnification in machine learning. Many machine learning models used with images today are scale invariant as they use a combination of convolutions and down-sampling to extract features at different scales. For your problem, as long as you ensure that your training set has plenty of examples at both 50x and 100x, you should be able to easily recognize features from both objectives without explicitly defining a 'pixel size'.
Examples of different objectives in the training set can be synthetically created during data augmentation by 'rescaling' images from other objectives (i.e. 50x and 100x). There are many tools under computer vision that will let you up/down sample an image to effectively 'change the pixel size'. Adding examples of 70x images (either real or synthetic) would be the best approach to ensure that features at this scale are learned.
Thanks for all the helpful answers and suggestions. Perhaps another way to achieve the goal would be to formulate my features with 'magnification' in mind before feeding it for training, so that the features are in units of micrometres instead of pixels. Would that be effective?
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training