Thread by @iScienceLuvr

Saw few tweets on pigeon-based classification of breast cancer (@tunguz @hardmaru, @Dominic1King, & ML Reddit), which was published in 2015. I work with the legend himself @rml52! I thought for my 1st Twitter thread I'd go over the papers's main points & our current work! (1/11)

My PI often likes to say AI stands for avian intelligence. And indeed his paper shows pigeons can learn the difficult task of classifying the presence of breast cancer in histopathological images. (2/11)

The pigeons were placed in an apparatus and the 🔬 image was shown to the pigeons on a touchscreen. The pigeons were given food if they pressed the correct button on the screen. (This is opposed to regular pathologists who are not given free food when analyzing images!) (3/11)

The pigeons were trained on 144 images (4x magnification) over a course of 15 days. Accuracy started at 50% (random) and ended up at 85%. Interestingly, performance on rotated and unseen test images was high, indicating generalizability. (4/11)

To understand what kinds of features the pigeons could be using, the pigeons were also trained on monochrome images (no color info) and performed decently as well. This could indicate that pigeons rely on textural cues, like some common deep learning algorithms as well. (5/11)

Finally, the commonly used ML technique of ensembling also works with pigeons! The scores of four pigeons were ensembles, obtaining an AUC of 0.99! (6/11)

Surprisingly, a lot of the images the pigeons failed at were truly difficult images that even pathology residents would struggle with. In particular, they failed at cases where the cancerous tissue had similar characteristics as normal tissue, and vice versa. (7/11)

There were some experiments with training pigeons to diagnose from mammograms, but training took much longer and the pigeons failed on the test set, indicating the pigeons probably were only able to memorize the images. (8/11)

In conclusion, pigeons aren't going to replace pathologists anytime soon but perhaps they are a good baseline to compare ML algorithms too 🤔
(9/11)

While it was a fun experiment that @rml52 loves to highlight to newcomers of the field/lab, we currentlywork on something a little different.
Our lab is instead focused on developing novel microscopy technologies like MUSE (@natBME). (10/11)

I focus on applying deep learning to both analyze and enhance images obtained from our 🔬 with the goal of hopefully integrating these technologies into a clinical pathologist's workflow. (11/11)

PS. I almost forgot to link the paper!

Also want to clarify this isn't my work, but my PhD advisor's work published 5 years ago. But I still thought it is very interesting work and worth sharing!

journals.plos.org/plosone/articl…

cc:
@jeremyphoward this thread might be of interest to you.
@anantm you probably already know about this, but thought I'd still share.