Completed on 6 Oct 2017 by Steven Hart. Sourced from https://www.biorxiv.org/content/early/2017/10/02/197517.
Login to endorse this review.
Really good walk through, including the history of CNNs. I'll make sure to require all my students to read this before working on one of these projects!
I was however left with a couple of questions/ suggestions that I thought might add some value if you're interested in feedback.
1. Methods state that whole slide images were only 2092x975 pixels. That seems way too small. Generally they are GB in size. Similarly, the algorithm was run on 20x20 pixels for basic-CNN and defaults for inception. Did you use 'tf.resize' for this?
2. What were the total number of image patches evaluated for each study? That seems like a reasonable question since you are actually assessing patches instead of slides, right?
3. When you say 'detect various cancer types by about 100% accuracy' , I think this is a little misleading. You are actually differentiating between lung, breast, and bladder tissue rather than cancer. That is, unless you are differentiating normal from cancer - but I don't see those results here.
4. You might want to expand a bit more on what it is you are classifying with the IHC. Are you predicting what marker it is or the scoring of that marker? Some of the text makes it read as if the former, while Figure 2 suggests it's the latter.
Also, work from the Google team showed improvement of patch classification by rotating the image multiple times and then averaging the logits to get a more robust measurement. I'd be happy to contribute the TF-Slim code I have for this (it fits in your train_image_classifier.py script).