Diagnostic Imaging AI — The Real Power and the Pitfalls of Reading Support— Sensitivity and specificity, the weight of a false negative, approval versus operation

01How Diagnostic Imaging AI Works — What It Learns, and What It Outputs

At the heart of diagnostic imaging AI is deep learning (= a kind of machine learning that automatically learns features from large volumes of data), and in particular the convolutional neural network (= CNN, a design that scans an image through small windows to pick up features piece by piece), which is well suited to images. It learns by matching tens of thousands to hundreds of thousands of images against the "correct answer" attached to each one (= whether this image contains a lesion, and where it is).

The key point is that what the AI learns is not "disease itself" but the patterns common to images that carry a correct label. Because it takes the diagnoses made by radiologists and pathologists as its model, the ceiling of the AI is basically set by "the quality of the diagnoses made by those doctors." If the model is biased, the AI inherits that bias.

The form of the output also varies by product. Some return "abnormality suspected / not suspected," some mark the location of a lesion with a box or color, some give the probability of malignancy as a number — the design diverges by use. What is common is that AI returns a probability; it does not make a definitive assertion. Miss this single point and the whole discussion of sensitivity and specificity that follows falls apart.

02Sensitivity and Specificity — Two Kinds of "Correctness" That Are Hard to Have Together

When we talk about the performance of diagnostic imaging AI, the two most fundamental measures are sensitivity and specificity. The words sound technical, but the substance is simple.

The advertising line "the AI is 95% accurate" often blurs these two. Sensitivity of 95% and specificity of 95% mean entirely different things, and in the first place, how many people in the target population actually have the disease (= prevalence) changes the real hit-and-miss outcome dramatically even at the same sensitivity and specificity. In low-prevalence screening, even a slightly lower specificity can produce a flood of false positives. When you look at a number, you must always check as far as "what was measured, in which population."

03The Weight of a False Negative — Why a Miss Is Especially Heavy

There are two kinds of error: the false positive, calling a healthy person diseased, and the false negative, missing a person who has the disease. In medicine, the weight of these two is entirely different.

A false negative is heavy because the chance to notice the error is lost. With a false positive, it often ends with "we took a closer look just in case and there was no problem." But with a false negative, the flow of testing stops the moment "no abnormality" is stated, and time passes with no one noticing the mistake. That is why diagnostic imaging AI, especially the kind used in screening, is often designed for high sensitivity even at the cost of accepting some false positives.

04Approved Cases — What Has Been Recognized, and How

Diagnostic imaging AI is no longer a fantasy. In Japan, under the framework of the program medical device (= SaMD, software approved or certified as a medical device in its own right) — the framework under which software including AI is treated as a medical device — products have been approved and certified that support the detection of diabetic retinopathy, polyps in colonoscopy, aneurysms on brain MRI, and more. The U.S. FDA has also authorized AI/machine-learning-enabled medical devices in the hundreds, many of them in the radiology, ophthalmology, and cardiology imaging fields.

What must be grasped here is the scope that "approved" implies. Approval is a limited permission that says "this product may be used, in this way (= intended use), for this target." An AI approved for judging diabetic retinopathy cannot be diverted to judging a different eye disease. Using it outside the intended use is use beyond the scope of approval. Where this distinction bites in pharmaceutical practice is, for example, when a diagnostic AI is discussed together with one's own product. Suggest efficacy even a single step beyond the approved intended use, and you touch the wall of the Pharmaceutical and Medical Devices Act discussed below.

05Operational Cautions — The Gap Between Approval and the Field

Approved performance is, in the end, performance on the data submitted for review. Bring it into the field and the same numbers do not necessarily appear. Several factors create the gap.

In particular, automation bias is easily overlooked. Once AI is introduced, people unconsciously trust it and their own attention to verify thins out. Then people can no longer catch the false negatives the AI missed, and the double net shrinks to a single one. Unless you set an operational rule at the time of introduction — "even a negative from the AI is independently confirmed by a person" — you can even end up with the absurd result that misses increase after introducing AI.

06The Relationship With the Physician — Where Does Responsibility Remain

"If AI does the diagnosis, who bears responsibility when it is wrong?" — this question always comes up in the field. The answer under the current framework is clear. The responsibility for the final diagnosis lies with the physician. One reason most approved diagnostic imaging AI is positioned as "support" is precisely to keep this locus of responsibility from moving.

So the desirable relationship between physician and AI has a clear order of precedence. AI is a tool that raises candidates, prompts attention, and reduces oversights. The physician is the subject who takes that as reference and judges comprehensively, together with the patient's background and other findings. The framing is that the AI's output is "one opinion," not "the conclusion." When this relationship breaks down and the AI's output is taken directly as the conclusion, it becomes a breeding ground for automation bias, and the locus of responsibility grows blurred as well.

When speaking to this point from a position connected to pharma, it is essential to stay with neutral fact. Elevating a particular diagnostic AI product with a tone like "more accurate than a doctor" or "with this you can rest easy" can be received as promotion or endorsement. It is safest to state only, as fact, that "it is a support tool and the final judgment rests with the physician," and to avoid ranking products by name.

07Verification — On What Basis Can We Say It "Works"

Whether a diagnostic imaging AI truly works is decided not by advertising numbers but by the quality of verification. Trustworthy verification has several conditions.

• Measured on data separate from the training data (= external validation) — Scoring well on the images used for training is a given. Only when performance holds on data from an entirely different facility and a different population can we call it real ability.
• Prospective clinical evaluation — Not merely analyzing past images in a batch after the fact, but examining, within the actual flow of care, how misses and diagnoses change with and without the AI.
• Sensitivity and specificity shown with the target population — Look for whether it is shown not as "95% accurate" but as far as "in which population, at what prevalence, what are the sensitivity and specificity."
• Who reviewed it — Whether it is reported in a peer-reviewed paper (= peer-reviewed, examined in advance by independent expert third parties). Do not swallow whole the numbers from a manufacturer's own announcement.

It matters greatly that a pharmaceutical medical affairs function carries this verifying eye. The perspective for evaluating a diagnostic AI's performance is continuous with the perspective for reading one's own clinical data. "In which population, compared with what, verified by whom" — the habit of asking these three is the foundation for not being swayed by AI's numbers.

08Connections to Other Chapters on This Site

This installment gains depth when read together with the following chapters and areas.

• AI Medical Vol. 4 — Electronic Health Records and AI — After the "seeing" information of images, how AI handles the "writing and keeping" information of records, summaries, and voice input.
• AI Marketing Vol. 5 — AI-Generated Content Strategy — Designing to deliver AI's output to the field while keeping regulation. The idea of leaving the final judgment to a person runs through it, just as with the "support" role of diagnostic AI.
• Material Review series — The practical safeguard of the discipline of not speaking of efficacy beyond the scope of the approved intended use.

Diagnostic imaging AI is the most fully implemented AI in medicine. Approved cases have grown, and it genuinely helps as a double net that reduces oversights. But its real power can only be spoken of once you correctly understand these things: the two hard-to-reconcile forms of correctness that are sensitivity and specificity, the especially heavy error that is a false negative, and the gap between approval data and the field. AI is a support tool that returns a probability, and the final judgment and responsibility remain with the physician. Do not be carried along by automation bias; have a person independently confirm even the AI's negatives. Only with this discipline does AI become an ally that reduces oversights.

From a position connected to pharma, there is one more fence. Do not suggest efficacy beyond the scope of the approved intended use, and do not elevate a particular product in a form that can be read as promotion or endorsement. When speaking of a diagnostic AI's performance too, stay with neutral fact. Next time we move from the "seeing" information of images to the "writing and keeping" information of the electronic health record — how AI is changing the clinical record.