Dr. Bibbo's Presidential Address on Automation in Cytology: Were Her Predictions Right, Wrong, or Somewhere in the Middle?

1983: Ronald Reagan was President of the USA, the last of the Delorian cars was produced (remember Back to the Future), Sally Ride became the first woman in space on space shuttle Challenger's second flight, Korean Air Flight 007 was shot down by Russian fighters after straying into Soviet airspace, the USA invaded the Caribbean Island of Grenada to rescue US medical students, catastrophic earthquakes in Turkey and Japan killed thousands, Martin Luther King Day was declared a US national holiday, hundreds of US Marines were killed in bombings in Lebanon, and I was just finishing my third year of pathology residency.

In another historical event (from a cytologic viewpoint) in November of that year, Dr. Marluce Bibbo gave the Presidential Address at the American Society of Cytology (now Cytopathology) in Denver, CO, USA [1]. The topic of her address was “Analytic and Quantitative Cytology” which, in 1983, was a 30-year-old field which was both maturing and at the same time approaching a philosophic junction. Dr. Bibbo, through her work at the University of Chicago with her mentor and associate George Wied, was already a recognized leader in this arena, and her address detailed the relevant background as well as her musings on the future and the way forward.

For readers who might not be so familiar with Dr. Bibbo's biography, a few “tidbits” are in order. Of course, all cytologists will know her from her textbook Comprehensive Cytopathology [2], now in its fourth edition, which is appropriately, and perhaps even affectionately, referred to by students merely as “Bibbo.” Dr. Bibbo is originally from Brazil where she completed medical school and her pathology residency in Sao Paulo. Coming to the USA in 1969, she completed a Cytopathology Fellowship at the University of Chicago where she remained on the staff, rising to the rank of full professor. She moved to, and is currently still active at, the Thomas Jefferson University in Philadelphia, where she is the Warren G. Lang Professor of Pathology and Cell Biology. She was Editor-in-Chief of Acta Cytologica from 2004 until 2013 [3].

Automated analysis of cervical cytology specimens as a screening method for cervical neoplasia had been the “holy grail” of a large number of investigations with large amounts of cash spent for the previous 30 or so years. Flow cytometric and image-based instruments had been developed which could recognize very fine differences in cell populations [4,5]. Anyone training the uninitiated first-time student of cytology will appreciate the subtleties inherent in this endeavor, even when working with a human brain! However, to date, these devices had not achieved usability in a practical sense - their cost was too high and their reliability suspect. Dr. Bibbo does, however, note that “considerable progress has been made” but laments on whether “fully automated cytologic diagnoses ... will ever be achieved” [1].

One might drop the subject right there - money down the rabbit hole - cut your losses - don't throw good money after bad - all of these are potentially appropriate comments, given the status of the field as described in 1983. But the problems to be addressed still existed:

1. Cervical cancer is preventable by screening (and is therefore desirable to accomplish). True.

2. Screening for cervical cancer by human eyes through the microscope is “tedious and exhausting”, and is fraught with at least a 10% false negative rate for missed cancer or precancer. True.

3. If all eligible women were screened by the current standard of manual human microscopy, the workforce is inadequate. True.

4. Improved Pap test preparation could make the test more accurate (but Dr. Bibbo opines that doing so would ultimately not be cost-effective). See below.

5. The most common cause of a missed cervical cancer is no Pap test at all. True (and No. 3. above becomes a real concern).

The point is, that despite the apparent dismal future for cytology automation, there is still a great need for it. So what does Dr. Bibbo set as the “state of the art” in analytic and quantitative cytology in 1983? And how should the field find fertile ground to overcome its inadequacies - she has a plan!

As was mentioned above, in 1983, the field was at a philosophic junction in the road. One of the paths represented what, in retrospect, was an error in practical thinking, i.e., that human observers do not have “planned” error rates. Early investigators were wedded to the principle that automated devices should not have an inherent (known) error rate; they should “be like humans” who never planned to make a mistake. Of course, we all know (but perhaps are not willing to admit [or quantify]) that humans do indeed have false-positive and false-negative rates and that these rates vary, depending on the process of thresholding. For example, some days I am more sensitive, perhaps because yesterday I found that I had missed a case via the retrospective review process. But inherent in this increased sensitivity is poorer specificity because today (in order to be more sensitive) I will have a tendency to overcall. If this sounds like the accuracy calculations we make for instrumentation in the clinical laboratory - well, it is indeed like that. Humans are not perfect and they have operating characteristics that vary just like machines do. However, leading up to 1983, cytologists were not willing to admit such, and hence they were also not willing to allow a device to perform autonomous clinical work with a “known” accuracy of less than 100%. It was thought that the use of such a device would be unethical when compared to the alternative of “no planned inaccuracies” in a human manual examination. The devices that had been developed had inherent (known because of extensive testing) disease detection sensitivities of 90-95%. In contrast, specificity, as Dr. Bibbo points out, was also a huge problem. The identification of what was truly normal was, in actuality, the most difficult task for instrumentation. Abnormal cells were actually much easier for the computers to identify, but if normal events were actually “detected” as abnormal (if the device were “set” to achieve maximal sensitivity), the number of “alarms” would easily overwhelm any automated system, essentially eliminating its practical use. When I entered the field in the late 1980s, this raging debate over the ethics of devices with known error rates being approved for use was stifling the field. As Dr. Bibbo notes, “It is unlikely, in our judgement, that any cytology smear reading device will be in clinical use in the near future” and that “FDA approval appears to be a long way off.” Fortunately, the junction in the road had another path that Dr. Bibbo alludes to in her presentation.

Strides in computer image analysis were advancing rapidly. New devices, such as laser scanning microscopes, were being developed that could acquire cellular images in high resolution from an entire Pap slide in under 1 min [6]. Computing power was lagging behind, so these microscopes had to actually be slowed so that the information would not overload the buffers. In the University of Chicago Laboratory, Drs. Wied and Bibbo could now image entire slides, storing the coordinates for potentially abnormal cells, which could then be later examined manually by human observers [7]. With such systems, the power and utility of the alternate pathway was about to be recognized. Perhaps devices did not have to be perfect. Why not let the instruments do the heavy lifting of prescreening and reserve further evaluation for humans? Let cytologists do what they do best: assess high probability “alarm” cells as to whether they are normal or abnormal. The concept of “guided-screening” functionality was coming into its own. This epiphany of thought moved the field down that alternative pathway. This was the turning point for automation. It still took another decade, but freed from the need for perfection, realistic and practical clinical applications were identified, computing power advanced by orders of magnitude, and commercial ventures sought and received FDA approval.

The industry was helped greatly by parallel developments in image acquisition and analysis. The development companies found unlikely bedfellows in defense departments. If satellites could obtain high-resolution images and image analysis could find tanks and missiles amongst all the camouflage, then similar analyses could power other medical applications in radiology and pathology. In fact, the first several commercial ventures in the automated cytology industry that achieved FDA approval (namely the PAPNET [8] and NeoPath AutoPap [9] systems, both approved in 1995) came from development teams originally involved in satellite image analysis. Interestingly, neither of these devices was initially approved for primary screening tasks. This reflected the need for this emerging technology to be heavily scrutinized by regulators. Both devices were initially approved for quality control rescreening, arguably a “hard nut to crack” in the commercial clinical arena. However, in laboratories where QC automation was implemented, small, but quantifiable and significant improvements were noted, namely in the reduced numbers of false-negative Pap tests [10,11]. Of course, the manufacturers had no intention of stopping with QC applications and they continued with FDA applications for primary screening - more robust economic business plans! This led to clinical trials which were subsequently successful for the AutoPap, first in 1998 as a slide profiler ranking slides based on probability of abnormality, and for which a percentage (25% in the USA, higher in the international market) could be archived as normal without any human review or further QC [12]; and secondly, in 2008 (renamed FocalPoint) as a true guided primary screening device, localizing potentially abnormal cells for further human scrutiny [13]. True to the discussion above, these devices were not perfect. The clinical trials of the AutoPap primary screener showed a 3% false-negative rate for HSIL [12]. There was also a fair degree of skepticism among users. Examples of device false-negatives could contain obviously abnormal cells that humans would “never miss.” However, the reality showed that the devices found more “subtle” abnormal cases that humans more frequently missed, and when the final scores were tallied, automated devices almost always found more abnormality than did human screeners, with the localization feature actually allowing users to more accurately classify the abnormal cells detected [14].

Dr. Bibbo also mentions the use of special preparations as a potential means to improve the automated screening process, but immediately dismisses this avenue as too expensive to be a likely pathway. As all readers will know, liquid-based cytology (LBC) preparations were introduced very shortly after her Presidential Address in 1983 [15]. I had the opportunity to visit the original Cytyc facility in the mid-1980s to review the new ThinPrep technology and the computer that was being designed from the ground up for screening the new LBC preparations [16]. As Dr. Bibbo mentions, DNA analysis was a potential method of abnormal cell detection, and the final version of the ThinPrep Imaging System (TIS) took advantage of this principle, using a new Feulgen-like stain which was stoichiometric for DNA (with an optical density proportional to DNA content), and could therefore identify cells having too much of it. This process took a significant amount of time, with the TIS not being finally approved until 2003, possibly not from a lack of technology, but because of another revelation that took place in the automated cytology industry. This revelation was a not-entirely-intended (but fortuitous) consequence of the investigation of the LBC method, as a “front end” to improve computer analytic performance. It was quickly found that it also improved the performance of human screeners. In addition, LBC had other advantages besides the improved cellular preservation and uniformity of appearance needed for automation. It randomized the sample, leading to improved abnormal cell “capture,” achieved significantly better adequacy rates, and observers found it a much more “pleasing” specimen to view. All of these benefits led to an improved Pap test, but as Dr. Bibbo noted, it would have been unlikely to improve the cost-benefit due to its much higher cost. The equation changed when FDA approvals were piggy-backed onto the cytologic examination for other tests from the same collected sample, most notably chlamydia/gonorrhea and HPV tests. As a “one-stop” test for a variety of important diseases, the case for using LBC and the cost-benefit equation changed. Laboratories very rapidly changed from using conventionally prepared slides to LBC. Now the common use of these special preparations could be much more easily and efficiently mated with automation: the best of both worlds.

So, as it turns out, was Dr. Bibbo right, wrong, or somewhere in the middle with regard to her predictions about cytology automation? Well, to me as an observer, it was a bit of all three. Her most important correct conjecture referenced the need to change the opinion about how automation would be used in the field of cervical cancer. Her observation that the devices would never be perfect and that there was a need to change that goal was clearly prescient, and presaged exactly what subsequently happened in the commercialization of automated primary screening. She was, however, wrong that it would be a “long way off” when, in fact, it was only a few years before QC devices which were also fully capable of performing guided primary screening were approved. History will certainly forgive this “inaccuracy” as the state of computerization and the tremendous improvement in image processing and analysis coming from other industries would have certainly been difficult to predict at the time. Her second “right” prediction was that special preparations might lead the way into automation. This prediction came true with a bit of a “side track” of human LBC screening, as its advantages for that task alone were appreciated. She was “in the middle” when it came to predictions about the use of nontraditional methods to detect abnormality. She spoke about “malignancy-associated changes (MAC)” as a means to screen (even in normal cells) for patients at a high risk of harboring cervical neoplasia [17]. The presence of MAC was a real finding, but its use was never shown to be of commercial value. She does not mention immunologic markers as a means to detect disease. In 1983, of course, pathologists were just beginning to use immunohistochemical markers as a method to identify “lineage” specific features in cells (e.g., keratin for epithelial cells, common leukocyte antigen (CD45) for lymphoid cells, etc.). Dr. Bibbo had no way of knowing at the time that markers of “transformation” (e.g., p16) or of specific genetic mutations (e.g., p53) would be available in the not-too-distant future.

So what does the future hold? In some ways, I currently share some of the uncertainty that Dr. Bibbo must have felt while giving her address in 1983. It is clear that the field of cervical cancer screening is still awash in its evolution. Widespread preventive HPV vaccination implementation will change the dynamic of how screening programs must be administered. Lower disease prevalence in the population is already being achieved [18] and, as a result, cytology screening sensitivity will also decline: the 2 factors (prevalence and sensitivity) go hand in hand [19]. Even automated screening, for which habituation (human inattention) has no effect, will see some decline in true-positive case detection (more high-scoring “alarm” cells will be mimics rather than real disease). Therefore, some countries have already moved to HPV primary screening with no cytology, or cytology only as a secondary test in some HPV-positive cases [20]. HPV molecular testing remains sensitive (either viral DNA is detected or it is not), but the specificity for preneoplastic disease is decreased (more cases will be infection rather than neoplasia, particularly in the vaccinated population). These problems, most of them caused by the success of vaccination, will require a more focused method of detection for the rare “breakthrough” true-positive cases at risk for cervical cancer development. The detection of molecular “events” which are specific to preneoplastic and neoplastic transformed cells may be the appropriate screening procedure in the future. Already, tests such as the colocalized combination of p16 and ki67, detected by dual-stain immunohistochemistry, have shown high sensitivity and specificity for high-grade lesions and cancer [21]. The best use of automated screening in the future may be the detection of this or another marker (e.g., methylation status) in cellular samples. Of course, such positive results could also be analyzed in solution without the need for any morphologic examination. In addition, the detection of specific genetic events tied to neoplasia could also be achieved by rapid (and increasingly cheap) screening genome-sequencing procedures.

Also on the horizon is the advancing integration of artificial intelligence in pathology. Cytology automation formed the basis for the current expansion of image analysis functionality into histopathologic automated specimen examination. Already, computer vision and deep learning (neural networks) have shown excellent results (sometimes better than humans) in specific tasks, such as finding small metastases in lymph nodes or identifying features, heretofore not identified, that predict outcome in cancer better than traditional morphologic grading systems [22]. Future practice - sometimes referred to as “computational pathology” - will involve the merging of many data streams [23]. For instance, cytology or histology features may be merged with radiology, laboratory, and/or clinical information in a complex matrix which might arrive at conclusions that are just not accessible to us with current methods.

In her closing paragraph, Dr. Bibbo opines: “We need not fear that automation will replace cytopathologists and cytotechnologists; instead, new developments will require acquisition of additional skills to keep up with modern systems …” This statement reflects the balance between comfort in our current positions and tension about events in the future, and is as true today as it was at the time of Dr. Bibbo's lecture. In 1983, the fear was for the replacement of humans by a machine that would perform the equivalent of manual screening. Today, it is for our replacement by technology that will require no manual screening or review at all. Cytotechnologists clearly see this “writing on the wall” as they have already begun the planning of an expansion of their scope of practice as so-called “mid-level practitioners” trained for a variety of morphologic-based tasks hitherto performed by pathologists [24]. This dovetails nicely with the already expanding role of the pathologist into the performance and interpretation of the many new tests coming on-line in immunology and molecular genetics. Clearly, the pathologist workforce is not expanding rapidly enough to perform all the newly generated work as well as the old routine tasks [25,26]. New midlevel practitioners should be ready to fill these gaps as they arise. Indeed, we do not have to fear for our jobs, we have only to be excited about all the new challenges that lie ahead. I am certain that Dr. Bibbo would agree!

The author has no conflicts of interest to declare.