Title and Logo

First published as:
Maverinck –
Why radiology must take care when it comes to AI.
26 September 2018.
Aunt Minnie Europe


This version has been extended with some additional references.
The version published in Aunt Minnie Europe was slightly shortened and edited to make it fit.


Rinckside
ISSN 2364-3889

Rinck PA.
Some reflections on artificial intelligence in medicine.
Rinckside 2018; 29,5: 11-13.
Read the Print Edition (PDF)



Some reflections on artificial intelligence in medicine

he point of artificial intelligence is that it “learns” on its own and becomes an – or even the one and only – expert. However, artificial intelligence is not as simple an approach as it's sold today, and artificial intelligence or expert systems are not recent ideas – they come and go since the 1940s, or even since the 18th century with Maelzel’s chess-playing automaton, The Turk.

The reliance on ad­vanced scientific theories and modes of reasoning and the utilization of scientific methodology, specifically observation, can easily lead to tunnel vision or wrong conclu­sions as it's known from the 19th century "ratiocination".

In 1843, the English philosopher John Stuart Mill differentiated in his book "A System of Logic, Ratiocinative and Inductive" induction from ratiocination, and developed prin­ciples of inductive reasoning:

"Reasoning, in the extended sense in which I use the term, and in which it is synony­mous with Inference, is popularly said to be of two kinds: reasoning from particulars to generals, and reasoning from generals to particulars; the former being called Induction, the latter Ratiocination or Syllogism … The meaning intended by these expressions is, that Induction is inferring a proposition from propositions less general than itself, and Ratiocination is inferring a proposition from propositions equally or more general [1]."

Two years earlier, Edgar Allan Poe described the same approach in his short story "The Murder in the Rue Morgue":

“But it is in matters beyond the limits of mere rule that the skill of the analyst is evinced. He makes, in silence, a host of observations and inferences. So, perhaps, do his companions; and the difference in the extent of the information obtained, lies not so much in the validity of the inference as in the quality of the observation. The necessary knowledge is that of what to observe [2].”


A hundred years later

A little more than a hundred years later, in 1958, the New York Times reported in an article that ...

“The Navy revealed the embryo of an electronic computer today that it expects will be able to walk, talk, see, write, reproduce itself and be conscious of its existence … The Navy said the Perceptron would be the first non-living mechanism 'capable of receiving, recognizing and identifying its surrounding without any human training or control.' The 'brain' is designed to remember images and information it has perceived itself … It is expected to be finished in about a year [3].”

It didn't work due to “technical limitations”.

spaceholder red600   The most famous first medical application of AI was MYCIN, a program developed in the 1970s at Stanford University in California [4].

MYCIN, as Bruce G. Buchanan and Edward H. Shortliffe described it in a recapitula­tion of the project, was a software that embodied some intelligence and provided data on the extent to which intelligent behavior could be programmed. The intention was to identify bacteria causing severe infections, such as bacteremia and meningitis, and to recommend antibiotics at the right dosage for a patient. As with other AI programs, its development was slow and not always in a forward direction.

It worked, but it also didn’t, and was never used in practice – not only because comput­ing power was insufficient, but rather for an inherent problem of AI: the knowledge of a human expert cannot be translated into digitizable rule bases. Additionally, AI is not im­mune against human prejudice that always exists – wittingly or unwittingly. Such pre­conceptions cannot be filtered out because of AI’s lack of a critical mind. Buchanan de­scribed this problem in a conclusion:

“There are many 'soft' or ill-structured domains, including medical diagnosis, in which formal algorithmic methods do not exist. In diagnostic tasks there are several sources of uncertainty besides the heuristic rules themselves. There are so-called clinical algo­rithms in medicine, but they do not carry the guarantees of correctness that characterize mathematical or computational algorithms. They are decision flow charts in which heuristics have been built into a branching logic [5].”


The flaws

AI is mindless, lacks consciousness and curiosity. These are fundamental flaws, distinguishing it from real intelligence. Although meant to be a “science” by its fa­thers, AI is not a real science; it’s closer to computer gambling and tinkering than to cre­ating a fundamentally reliable support system for highly specific tasks.

ruler black

Artificial intelligence is mindless. This is a fundamental flaw.

ruler black

Neural AI networks are good at – crudely – classifying pictures not only in radiology, meanwhile they encompass the entire spectrum of medical imaging, including for exam­ple nuclear medicine, dermatology, and microscopy. The are known for years as CAD, computer-assisted diagnosis.

spaceholder red600   A typical example is a recent paper by a dermatology group at Heidelberg University. They used deep learning neural networks for the detection of melanomas. The British newspaper The Guardian summarized the press release from Heidelberg with the head­line: “Computer learns to detect skin cancer more accurately than doctors”. The authors of the paper concluded: “Most dermatologists were outperformed by the neural net­works. Irrespective of any physicians' experience, they may benefit from assistance by a neural networks’ image classification [6].”

In an editorial accompanying the dermatology article in Annals of Oncology, the com­mentators were more careful and raised some additional concrete questions.

“This is the catch; for challenging lesions where machine-assisted diagnosis would be most useful, the reliability is lowest.” They also point out: “Whilst dermatology is a vis­ual specialty, it is also a tactile one. Subtle melanomas may become more apparent with touch as they feel firm or look shiny when stretched [7].”


Legal responsibility

Another main problem of AI is that the overwhelming majority of its users do not un­derstand and cannot follow its black-box judgments and its reasoning to reach certain choices. Interestingly, there also are a number of reports that developers of AI software did not understand why their algorithms reach certain results and decisions; the algo­rithms are impenetrable.

Thus, the well-meant “right to an explanation” of decisions made by an AI expert sys­tem concerning a person, passed as a European law in the General Data Protection Regu­lation (GDPR), can hardly be fulfilled because if even some creators are unable to find inherent flaws in their source code they won’t be able to explain it to their “vic­tims”. I wonder what the legal consequences will be.

It is a principle of information technology that convenience and security are generally mutually exclusive. Once again the question arises whether the limits of what is ethi­cally permissible are being shifted because something is technically possible. How­ever, financial and career interest often override established values of the medical pro­fession. More so, there are other interests in forcing the introduction of AI by groups and institu­tions owing no allegiance and acknowledging no responsibility to patients, doctors or the people in general.

At this point we are faced with another question – who is really responsible and accountable for the quality of the results? The radiologist, the hospital’s administrator, the software engineer who wrote the source code, the company that sold the software? The compa­nies will reject any responsibility, stating that the AI software was delivered free of de­fects. Even if the customer will get access to the source code, nobody will ever be able to prove that the algorithm has a flaw. You bought a pig in a poke – and are stuck with it.


Understanding AI

There are other problems. In a recent overview of AI in AME the author stated:

“The accuracy of these algorithms is dependent on two important factors: the type of al­gorithms used and also the acquisition parameters applied by the modality. If the algo­rithm is to be accurate, it is really important the acquisition parameters are standardized prior to application of the algorithm [8].”

This is a major dilemma of AI and deep learning. In many instances, the calculated pa­rameter data are incorrect, as we have seen in “MR fingerprinting” and related method­ologies. These values cannot be reliably reproduced, thus they shouldn’t be used in a neural network [9]. Deep learning can lead to the description of complex relationships that might only exist because they are based on artifacts or wrong presumptions.

Simple tasks are easily solved by AI, multi-layered tasks are far more complicated to be worked out. During the last ten years, neural networks have shown promises. Still, AI doesn’t mean an understanding, thinking, and comprehending computer, but pro­grammed if-then ordered decisions. At the present stage, artificial intelligence is more real incompetence that easily can run wild and lose control than helpful support in diag­nosis.

spaceholder red600   AI is also claimed to be objective. But there is no objectivity or neutrality in AI, its deci­sions are not necessarily knowledge based, but biased. More so, quantifying algorithms freeze a state of the past because they use old data.

Artificial imaging programs are useless if applied randomly without a well-defined and sharply delineated aim. Many approaches to explain results of AI are based on hypothe­ses which are still to be proved, and much research in this field is empirical and heuris­tic.

spaceholder red600   Still, AI will come on the market; it’s business value is enormous. By the way: If AI should work, even limping and stuttering, other disciplines will take over radiology in those fields which they find attractive – because with fast AI results it’s easy and makes money. Anyone can use it, from technologists to physicians in clinical disciplines. Radi­ologists are not needed for this.


spaceholder blue

References

1. Mill JS. Of inference, or reasoning, in general. In: A system of logic, ratiocinative and inductive, being a connected view of the principles of evidence, and the methods of sci­entific investigation. Volume I, London: John W. Parker Publisher. 1843. p. 223
2. Poe EA. The Murders in the Rue Morgue. Philadelphia (USA): Graham's Magazine 1841.
3. UPI. New Navy device learns by doing – Psychologist shows embryo of computer designed to read and grow wiser. New York Times. 7 July 1958. p. 25
4. Buchanan BG. A (very) brief history of artificial intelligence. AI Magazine 2005; 26,4: 53-60.
5. Buchanan BG, Shortliffe EH (eds). Rule-based expert systems: The MYCIN experi­ments of the Stanford Heuristic Programming Project. Reading, MA, U.S.A.: Addison Wesley. 1984. pg. 683.
6. Haenssle HA, Fink C, Schneiderbauer R, Toberer F, Buhl T, Blum A, Kalloo A, Has­sen ABH, Thomas L, Enk A, Uhlmann L; Reader study level-I and level-II Groups. Man against machine: diagnostic performance of a deep learning convolutional neural net­work for dermoscopic melanoma recognition in comparison to 58 dermatologists. An­nals of Oncology; 28 May 2018. doi:10.1093/annonc/mdy166
7. Mar VJ, Soyer HP. Artificial intelligence for melanoma diagnosis: How can we de­liver on the promise? Annals of Oncology; 28 May 2018. doi.org/10.1093/annonc/mdy193.
8. Rinck PA. Relaxation time measurements in medical diagnostics. In: Rinck PA. Mag­netic resonance in medicine. A critical introduction. 12th ed. BoD, Norderstedt, Ger­many. 2018. pp. 87-92.
9. Dugar N. AI algorithms begin to loom large in radiology. AME 27 June 2018.

spaceholder blue003 spaceholder blue003

VirtualCampus

00-f1

00-f2

00-f3

00-f4

00-f1

00-f2

00-f3

00-f4

00-f1

00-f2

00-f3

00-f4

00-f1

00-f2

00-f3

00-f4

00-f1

00-f2

00-f3