greenelab/deep-review

Deep convolutional neural networks for annotating gene expression patterns in the mouse brain

cgreene opened this issue · 2 comments

Converted Allen Brain Atlas -> present/absent calls for mouse images. Used OverFeat architecture. I'm pretty sure they used OverFeat trained on ImageNet images - especially since they resized to 231x231 to meet the requirements of the model. They also make some statements to this effect.

They then trained a classifier for each expression pattern using the features at a given level. How training/testing/eval are done is a bit unclear. They do something with thresholds for class balance: "That is, at a given ontology level, we randomly selected training samples from the data set and checked whether the ratios between two classes among all structures were above a certain threshold. We repeated this process for a maximum of 5,000 times and then decreased the threshold if the ratio was not satisfied. Thus, the final thresholds are different for different data sets."

Unfortunately this isn't talked about in the methods and there doesn't seem to be a supplement. The story is a interesting: "you can train on generic images and the same features are predictive of mouse brain IHC!" It's just hard to know what to do with the paper when the evaluation is described in this manner and there's no source or supplement available to figure out how it was done.

Might be worth mentioning as an example of generic features shared across systems, but if we use it we should keep the level of detail in mind.

Going to close this for now. Could go into some discussion of transfer learning There are other examples that are more explicitly transferring within bio domains - this one brings in features from pretty far afield.