Self-Driving Microscopes?

A recent AACR presentation and an “under review” manuscript by the Google AI Healthcare team presents a fascinating (if somewhat idiosyncratic) light microscope that incorporates built in real-time artificial intelligence capabilities.  The authors developed deep learning algorithms to identify metastatic breast cancer in lymph nodes and prostate cancer in prostatectomy slides.  As a pathologist views slides through this adapted microscope, the deep learning results are projected into the optics in near real time (see image from their manuscript).  Essentially, areas of cancer are outlined in bold colors, drawing the eye to these areas.  The authors report that: “Pathologists testing the device reported a seamless experience that provided immediately useful information.”Screen Shot 2018-05-30 at 3.50.37 PM 2

I must admit to some confusion about the intended use of this technology.    It seems likely that the adoption of digital pathology would surpass any realistic need to develop this light microscope-based technology.  The authors point to possible applications in small labs or developing countries.  Regardless, it is a harbinger of exciting things to come in the word of digital pathology and deep learning.

This integration of an “assist device” into the light microscope seems analogous to the Society of Automotive Engineers’ taxonomy for the development of self-driving cars (see below) which ranges from no autonomy (0) to full automation (5).  I would argue that most pathologists currently practice in a 0-1 range.  Don’t we often ask one another:  “would you like to drive?”  Assist devices that elevate our rank from 0 to 1 may include ancillary tests like immunohistochemistry.  The only pathology devices that actually replace some pathologist functions entirely may be emerging image analysis tools that quantitate Ki-67 and ER/PR IHC.  The much-feared full automation (5), and consequent replacement of the pathologist, seems far down the road, but a very exciting journey!



Humanistic Digital Medicine

I’m a fan of Abraham Verghese and his writing; however, I have to take issue with several of his comments in a recent (otherwise insightful) article in the NYT entitled “How tech can turn doctors into clerical workers.”  His main premise is that the electronic health record (EHR) and machine learning have not lived up to their potential for improvements in population health (probably true).  Instead, the EHR has become  a threat to basic clinical judgement and the cause of significant physician burnout.

.  ViviCam 6300

There is no doubt that completion of the EHR is a burden to many physicians.  The timely documentation of the physician-patient experience (and its accurate coding) seems more important than the interpersonal experience itself.  My colleagues in areas like general internal medicine and pediatrics refer to “pyjama time,” i.e. the two to three hours they spend at home, after their clinics, completing their medical records.  Thank goodness we’re pathologists!

One disturbing comment made by Verghese: “…as with any lab test, what AI will provide is at best a recommendation…that a physician using clinical judgement must decide how to apply.”  I must disagree.  A positive HIV test, for example, is more than a recommendation to treat the patient with antiviral drugs!  And of course the idea that the application of clinical judgement is the exclusive domain of the non-pathologist is absurd.  But this role in clinical judgement demands that pathologists continually educate themselves about clinical decision-making and therapeutic options.  (On the future role of AI in pathology, see my upcoming blog posts.)

What can pathologists learn from the EHR experience as we become more digital and more involved in machine learning?  Verghese states: “…the leading EHRs were never built with any understanding of the rituals of care or the user experience of physicians…”  We pathologists, particularly academic pathologists, must engage early with the developers of digital pathology interfaces (software, monitors, mice) to advise them about the user experience and ensure they enhance our abilities and don’t lead to frustrations and inefficiencies.

Another disparaging comment about pathology from Verghese:  “Caring is expressed in listening, in…the bedside exam, …not on a biopsy report.”  Some of the most caring and empathetic physicians I have ever known are pathologists.  That humanistic element of caring is an essential part of every good pathologist, and must be recognized if were are to become equal partners in the clinical care of our patients.

One Size Doesn’t Fit All: biomarker accuracy in small biopsies

It is increasingly obvious that tumors are highly heterogeneous at a molecular level, and this heterogeneity likely plays a critical role in their behavior and response to therapy.  Consequently, the issue of sample size (and the number of cells scored in an assay) is one of critical importance for cytological specimens, mainly due to the risk of false-negatives in under-sampled or under-scored samples. Unfortunately, sample size is rarely addressed in a rigorous and probabalistic manner when developing ancillary studies.

file2781242214133An interesting online tool has been created by Christoph Hafemeister in the research laboratory of Rahul Satija that theoretically calculates the minimal number of cells needed (at a desired confidence level, e.g. 95%) to detect rare cells at a given frequency (e.g. 1% of the population).  For example, in the case of a current ancillary test of importance to immunotherapy selection, PD-L1 immunohistochemistry in non-small cell lung carcinoma, a patient whose tumor displays no staining (<1%) is not eligible for pembrolizumab therapy, whereas a tumor that displays “any” staining is eligible for this therapy.  

I created the following table using the online tool, making the assumption that there are only two cell types of interest (biomarker positive and negative) and that detection of a single positive cell in a sample defines positivity.  Obviously, the percentage of target cells in a population is critical in determining the appropriate sample size.

Screen Shot 2018-02-25 at 10.45.33 AM

Interestingly, the ASCO/CAP HER2 testing guidelines recommend the scoring of just 20 cells in ISH assays.  The above analysis suggests that false-negative results may be produced if the HER2 amplified cells are present at less than 10% of the population.  This is particularly relevant to fine-needle aspiration biopsy samples of metastatic lesions where the sample size may be low.  Of course the biology and therapeutic response of tumors with rare biomarker-positive cells is always a question.

We are already in a more-with-less era, and expectations for biomarker assays on small core biopsies and FNAs continue to rise.  Image analysis and deep learning tools may help with our accuracy, but they will have to be trained by pathologists to ensure sample adequacy.

Blood, Sweat and Tears: the liquid biopsy penetrates early detection

Pathologists have obviously been involved in early detection of cancer for decades.  The Pap test and HPV testing have dramatically reduced cervical cancer incidence in the US and PSA blood testing has played a significant, if controversial, role in prostate cancer screening.  But lab-based early detection of many cancers, such as pancreatic and ovarian, has remained elusive, resulting in many patients presenting at advanced stages. Detection of circulating tumor cells and circulating DNA (ctDNA) has prompted excitement, and may still play important roles in therapeutic monitoring of advanced cancers, but their utility for localized cancers may be limited.

file000709986619However, a recent report from some of my previous colleagues at Johns Hopkins has demonstrated encouraging sensitivity of a multi-analyte blood test for the detection of several types of non-metastatic cancers.  The new blood test, termed CancerSEEK, utilizes combined assays for genetic alterations and protein biomarkers.  The genetic alterations are assessed by a sensitive multiplex-PCR using a 61-amplicon panel that includes somatic mutations in 16 genes (TP53, KRAS, PIK3CA, etc.) commonly altered in the target cancers (ovary, pancreas, liver, stomach, esophagus, colorectum, lung and breast).  In addition to ctDNA the assay also targeted 8 plasma proteins (CA-125, CEA, etc.) previously found to be important in cancer detection.

This assay was then tested in over 1000 patients with Stage I to III cancers.  Sensitivity of detection ranged from 98% for ovarian cancers to 33% for breast cancers.  The specificity was also high, with only 7 of 812 individuals without cancer scoring positive.  The strength of this assay likely lies in its unique combination (and algorithmic analysis) of protein and genetic biomarkers.  These investigators estimate the price of such a test to be approximately $500.  Of course one important caveat to this study is that these patients were already diagnosed with cancer, so expansion of clinical trials to completely asymptomatic patients will be important.  Also, the sensitivity for the ideal targets, the Stage I cancers, was significantly lower (43%) than higher stage tumors.

Enthusiasm for sensitive screening tests should be balanced by the recognition that they pose the risk of prompting unnecessary biopsies and creating concern among our patients.

Atlas Shrugged

Don’t worry, this isn’t a commentary on Ayn Rand or her political philosophy!   I actually want to discuss atlases, because I love maps, and because they seem to be in vogue among Big Science initiatives.  The dictionary actually has a very quaint definition of atlas:  “a bound collection of maps often including illustrations, informative tables, or textual matter.”  Seriously, a bound collection?  Have our lexicographers never seen the internet (the US Library of Congress has over 28,000 digitized maps online)?


More specifically, I’d like to discuss some of the emerging “atlas” initiatives to characterize biological material.  (We’ll get to the “shrug” later in the post.)  Below is a partial list of several current atlas projects:

Virtually all of these aim to construct comprehensive, high-resolution, multidimensional, multi parametric atlases of normal and diseased tissues.  All of these efforts are being driven by amazing technological advances that allow unprecedented, deep molecular characterization at the single cell level, either in dissociated tissues or in situ (which should excite anatomic pathologists!).  Some of these efforts, such as HuBMAP and HTA are sponsored by the NIH and actually have funding available for technology development, etc.  Others appear to be data sharing consortia or data repositories.

So, back to the shrug (a gesture of doubt). Several of these initiatives acknowledge other, related atlas programs and claim a desire to synergize with them.  Time will tell how collaborative these programs will be and how successful they will be at avoiding redundancies.

A quick perusal of the involved investigators reveals relatively few pathologists.  Hopefully some of us pathologists will be motivated to participate and provide expertise!

Entering the Blogosphere

This is the post excerpt.


Happy New Year and welcome to my first blog posting! Perhaps the initial question to address in one’s inaugural blog posting is: why create a blog? For me the answer lies in my efforts to make sense of the rapidly changing terrain in science and medicine, and its impact on the field of pathology.


My perspective

In this first blog posting it might be useful to provide you with my professional background so you can see what experiences shape my opinions. Most of my professional life has been spent at academic medical centers on the East coast of the US. During part of that time I was also the Scientific Founder of a small biotechnology company. My research interests have ranged from basic cell biology to translational molecular biology, particularly in the areas of oncology and molecular biomarkers. I am also a practicing diagnostic Anatomic pathologist, primarily in the area of Cytopathology. Four years ago I left the East coast to become the chair of the Pathology Department at the University of New Mexico.


One way to examine the changes taking place in pathology is through the lens of disruption. The dictionary defines disrupt as: “to break apart or throw into disorder.” However, it may also mean:  “to interrupt the normal course.”   There are several disrupters that will have large impacts on the normal course of pathology that will likely become the topics of my future blog posts.

Molecular technologies and big data

We’re rapidly approaching the point in a research setting where we can identify virtually every gene, mRNA and protein in a population of cells. In fact, we’re getting closer to be able to deeply characterize every single cell in a population of cells. Our clinical challenge is to identify which of these has clinical validity and to generate assays that are analytically valid. To separate the signals from the noise in these data will require robust bioinformatics tools and expertise.
Another source of big data lies in the information we generate every day in the clinical labs. It is likely that there are valuable clinical data buried within the trends and patterns of individual laboratory data points that can be identified and better utilized for proactive health interventions. This too will require expertise in informatics, computer science, and even mathematics.

Computational pathology

Computational Pathology encompasses several different areas, including digital pathology, image analysis, machine learning and deep learning. Digitizing Pathology images will be the foundation for development of sophisticated image analysis tools and even deep learning assist devices to aid in location and interpretation of pathological entities.


We must acknowledge that declining reimbursement trends will stress traditional pathology labs (particularly those not functioning at peak efficiency). While we must strive for fair reimbursement, we must simultaneously ensure that we are key players in areas where reimbursement will be focused in the future, including overall clinical quality, safety, value and outcomes.


Although these disrupters will create some growing pains, and will require an infusion of new skills into our profession, I see tremendous opportunities to transform pathology into a vibrant and critical profession in the precision medicine of the future.  Hopefully my future blogs will encourage a dialog that will inspire us on this path. While I will strive to make my posts evidence-based, they will be filled with my opinions and are not intended to be comprehensive reviews of the topics I discuss.
Until next time!