Then I saw their AI – Now I’m a Believer

Not a Sigh of Doubt in Deep Mind…(to the tune of “I’m a believer” by Neil Diamond with apologies)

Like most pathologists, I am skeptic – a real Doubting Dave. I have seen a lot of fads come and go in medicine and pathology. I thought that Artificial Intelligence (AI), in anatomic pathology was another fad. In fact, as Dr. Toby Cornish from Colorado University in his talk, “Artificial Intelligence in pathology” for The Pathologist pointed out, we are at the peak of the Hype Curve. It will probably be several years before we reach the “slope of enlightenment” and the “plateau of productivity”.

However, 2017-2018 seemed to be a breakout time for AI in anatomic pathology and events have changed my mind. The first was a rush of articles showing that AI algorithms, usually deep learning or convolutional neural network algorithms could 1) accurately identify metastases to sentinel lymph nodes in patients with breast cancer. 2) Some algorithms were superior to pathologists in detecting metastases, if the pathologist were under a time pressure. 3) The combination of the best AI and a pathologist, not under time pressure, resulted in the greatest sensitivity and specificity in detecting metastases. 4) AI could more accurately assign non-small cell carcinoma of the lung (NSCCL) to squamous or adenocarcinoma than pathologists. 5) AI could accurately assign Gleason Grade on prostate biopsies and prostatectomies. I recommend the article “AI-based Breast Cancer Nodal Metastasis Detection: Insights into the black box for pathologists” in Archives of Pathology and Laboratory by Liu, et. al. to get a better idea of how image AI is working.

The second event that got my attention, although this contributes to The Hype, was the number of anatomic pathology Startups dedicated to Artificial Intelligence and the amount of venture capital money flowing to these startups. I will blog about these companies soon.

What has changed? Why is AI in Pathology happening now? One big change is the increasing use of digital pathology. The advent of whole slide imaging has led to large amounts of image data that is now available to be mined. For example, there are large sets of whole slide images of tumors publicly available from The Cancer Genome Atlas (TCGA). If pathology remains analog, i.e. glass slides, AI would not be possible. Another big change is the big investments in AI made by the FAANG companies (Facebook, Apple, Amazon, Netflix and Google) are paying off and are spawning many AI spinoffs. Yet another change is that Convolutional Neural Networks, although not new, have been found to work especially well with large image data sets, e.g. Facebook’s facial recognition AI. (The word “convolution” is a mathematical term for “a function derived from two given functions by integration that expresses how the shape of one is modified by the other” and does not refer to the twists and coils on the surface of the brain.) Another big change is the improvement in Graphic Processing Units (GPUs) created for video games that are much faster at processing image data than Central Processing Units (CPUs). Other big changes that have supported AI in pathology were the development of the software that enables artificial neural networks to run on distributed cloud platforms like Amazon Web Services which is much more efficient, as well as the availability of open-source libraries of AI code and algorithms, such as TensorFlow. These two developments allow relatively small startup companies and academic pathology departments to develop pathology AI.

While we may be at the peak of a hype cycle, “I’m a believer” that AI is eventually going to profoundly change pathology.

Photo attribution: Liu, Y. et. al., Arch Pathol Lab Med. doi: 10.5858/arpa.2018-0147

Cornish, T., “The role of artificial intelligence in digital pathology,” Webinar sponsored by The Pathologist.