This weekend, I'm reading Practical Dermatopathology by Ronald P Rapini. It seems a very well organized book. There is an interesting analogy in the book. It says:
Pathologists may be subcategorized into home-run hitters and hedgers. The home-run hitters try to "force" a diagnosis, and give only one most likely diagnosis. They are either very, very correct, and look very smart, or else they strike out and miss the diagnosis completely. This can be dangerous. For example, they might diagnose a lesion as a definite Spitz nevus, which subsequently is found to be a melanoma when it metastasizes. Most Spitz nevi can be diagnosed with relative certainly, but there are always those difficult cases for which all the experts can have their opinions, using the best of criteria, but for which there remains an element of uncertainly. Simple histology has its limits in predicting biologic behavior. Hedger pathologists, by contrast, seldom make a specific diagnosis, and instead often give a long differential diagnostic list, even to the point of listing histologic possibilities that are ridiculous from a clinical standpoint. They rarely strike out, but they are sometimes not very helpful, and are not appreciated by clinicians. Wise pathologists avoid these two extremes.This reminds me of a concept brought up by a pathology resident, the idea of spotters or 'instant diagnoses' of entities which are instantly recognizable, versus other types of cases that can be more complex. I suppose the home-run approach would work great with spotters, whereas hedging would be most applicable to more ambiguous situations. Another resident told me, 'the best pathologists consider differentials before honing in on a diagnosis'. Essentially, going from good to great requires finding a happy medium. "The art of pathology is to be dogmatic about the diagnosis as often as possible, while not being afraid to hedge and give a differential diagnosis when the diagnosis is uncertain."
Today, I want to discuss the idea of pathology as art, within medicine, involving subjectivity. The reason for this is because of the strange idea that technology, such as high-powered computers or robots, will someday replace pathologists.
People who say pathologists will disappear as a specialty under the threat of AI do not really understand pathology. When this topic inevitably comes up in clinic, and I try to explain why AI would not be a threat to pathologists, it seems the clinicians become disinterested in the conversation. Why this is the case, I am not really sure, since most of the time they were the ones who brought it up as a topic of discussion in the first place, in response to me telling them that I am interested in going into pathology.
Thus, I want to list my personal opinion regarding why I think AI will not threaten the livelihood of pathologists everywhere.
1) Big tumors and sampling - I watched a TEDx talk where the presenter showed machine learning in which the computer was fed images of biopsies, and was able to stratify them in terms of prognosis with some accuracy. There are many judgements which a pathologist must make regarding biopsies and large cases, which a machine is incapable of. A machine is capable of producing output based on the data given to it previously, however, would it be able to determine whether a biopsy is adequate (in other words, representative of the lesion) [absence of data], or when it comes to larger specimens, whether the tumor has been adequately sampled during grossing? For some reason, I find it hard to believe that a machine would be able to determine whether additional sections of a large tumor need to be put in. These are judgements which an experienced pathologist makes easily and instantly when assessing a case.
2) Ontology and context - a machine does not understand the true meaning of what it sees. Yes, a person (i.e. computer scientist) can assign different values to data within the program/algorithm. However, truly understanding disease processes is something I cannot see a machine doing now or in the future. For example, a pathologist notices that a prior biopsy or sample had changes similar to the one he/she is noticing today, and based on the patient's lab tests and chart, concludes that the patient has a particular disease. The pathologist is able come to this conclusion based on a compilation of disparate sources of information, observation, analysis, and medical knowledge.
This brings me to the next two points.
3) Stains and artifacts - A machine (read, computer) does not know what is plausible or implausible. Preparation artifacts occur on a daily basis. Yes, a machine could be taught to recognize these, but there may be misinterpretations because of the variable presentation of such artifacts, given that they may range to subtle to overt. This highlights the often undervalued role of the pathologist as serving in QC (quality control) of the histology lab where slides are prepared. Pathologists are the arbiter of what is allowable in the quality of histology slides that are made for diagnostic purposes, for only pathologists have the privilege of conferring diagnostic meaning to a slide. Moreover, there are many mimics in pathology. Would a machine be sophisticated enough to distinguish these. Lastly, stains which require the use of a polarized microscope/direct immunofluorescence (DIF), I see a machine struggling to interpret given that it is based on light pattern and intensity.
4) Communication - One of the jobs of a pathologist is to serve as a liaison between clinician and the laboratory given the clinical context, in instances calling the clinician to get more information or recommending the appropriate ancillary testing, or simply communicating results in tumor boards, etc. Appropriate communication is something machines still have not mastered, just note our communication frustration with Alexa, Siri, and Google Home and various other voice activated/dependent products on the market currently. Formulation of reports are also important methods of communicating to clinicians in a concise, accurate fashion our assessment of a particular case. I feel doubtful that a computer or machine would be able to fashion a report which serves to be a useful consult to a clinician.
From my experience in pathology thus far, I feel AI will be a helpful adjunct, which may increase the efficiency of existing pathologists. Where I do see AI playing a role, is in removing some of the tediousness of pathology, and introducing more standardization, for example in the interpretation of stains such as HER-2 IHC, where there is interobserver variability. Just like AI was slotted to make radiologists "extinct" years ago, when radiology was brought from the 'dark ages' of silver-stained films to digitalized imaging and reading rooms, pathology is similarly on the cusp of making the transition from physical microscopes and slides to completely digitized workstations. Just see how they do it in Belgium.
One of the revolutions in pathology which I think will happen in my lifetime is the advent of telepathology, in which pathologists are freed to live and work in whichever location that they choose while reading virtual slides for a laboratory based in another location. This may help alleviate trainees' concern in pursuing a specialization which may not be particularly in demand, or concerns about finding a suitable position in a desirable location. Overall, the advances of modern day (tech revolution) hold a lot of promise as a tool for education as well, and I am excited to see how far the next generation of pathologists take this.
No comments:
Post a Comment