# AI in Diagnostics and Medical Imaging
A stroke patient arrives in the emergency department. A CT scan is taken at 2:14 a.m. Normally it sits in a queue behind twelve other scans until a radiologist gets to it. But an AI triage tool flags a suspected large vessel occlusion (a blockage in a major brain artery) and pushes the case to the top of the list. The neurologist is paged. Treatment starts 40 minutes sooner.
That is the promise. Now let us dissect why most imaging AI does not deliver it, and what separates the tools that improve outcomes from the ones that just generate alerts.
Start with the vocabulary, because it drives everything.
An imaging AI does not "read" a scan the way a radiologist does. Most cleared tools do one narrow job: detect the probability that a specific finding is present. Large vessel occlusion. Intracranial hemorrhage. A suspicious lung nodule. A collapsed lung (pneumothorax).
The output is usually a number between 0 and 1: the model's confidence. You then pick a threshold. Above it, the case is flagged. Below it, nothing happens.
That single choice, where to set the threshold, is the entire business and clinical decision. Everything else is downstream.
Two terms define diagnostic performance. Learn them cold.
You cannot maximize both. Lower the threshold to catch more disease, and you catch more healthy patients too. Raise it to reduce false alarms, and you miss real cases.
For a stroke triage tool, you tune for high sensitivity. Missing a treatable stroke is catastrophic; a few extra false alarms are tolerable because a human reviews the scan anyway.
For a tool that autonomously clears normal scans with no human review, you need extremely high specificity and sensitivity both, because a miss goes uncaught.
The Google ML crash course on classification thresholds walks through this tradeoff with clear visuals if you want the mechanics.
Consider a chest X-ray tool flagging possible pneumothorax in an ICU.
Set sensitivity too high, and radiologists get flooded with false alarms. Within a week they start ignoring the alerts. This is alert fatigue, and it silently kills the value of the tool. The AI can be technically excellent and clinically useless at the same time.
The right threshold depends on disease prevalence, staffing, and consequence of a miss. It is a workflow decision dressed up as a math problem.
An AI model is only as good as the "correct answers" it learned from. In imaging, that answer is called the ground-truth label, and getting it right is harder than people expect.
Who decides a scan truly shows a hemorrhage?
If the labels are noisy, the model learns noise. If the labels come mostly from one hospital's older scanner, the model learns that scanner.
Here is a concrete failure mode. A model is trained on scans from large academic hospitals using a particular CT machine and imaging protocol. It performs beautifully in validation.
Then it is deployed at a rural community hospital with a different scanner, different slice thickness, a sicker or healthier patient mix. Performance drops. Nobody notices immediately, because there is no radiologist double-checking every negative.
This is distribution shift: the real-world datareal-world dataRWD, données collectées en dehors des essais cliniques contrôlés : dossiers médicaux, claims d'assurance, données de dispositifs connectés, base des Real-World Evidence (RWE). differs from the training data. It is the single most common reason imaging AI underperforms after deployment. Regulators increasingly expect monitoring for exactly this.
In the United States, most imaging AI reaches market through the FDA's 510(k) pathway, which clears a device by showing it is "substantially equivalent" to an existing one. Important nuance: 510(kkThe average number of new users each existing user generates through referrals. Above 1.0, growth compounds on itself and becomes exponential.Voir la définition complète →) clearance is not proof the tool improves patient outcomes. It shows the device performs as intended on a defined task.
So when a vendor says "FDA cleared," ask the harder questions.
Questions to ask any vendor:
You do not need to build a model to understand the lever. This snippet shows how one threshold produces very different behavior.
# Model outputs a probability of "large vessel occlusion" per scan
scores = [0.05, 0.22, 0.61, 0.78, 0.91] # 5 scans
truth = [0, 0, 1, 0, 1] # 1 = disease truly present
for threshold in [0.5, 0.8]:
flagged = [s >= threshold for s in scores]
tp = sum(f and t for f, t in zip(flagged, truth))
fp = sum(f and not t for f, t in zip(flagged, truth))
fn = sum(not f and t for f, t in zip(flagged, truth))
print(f"threshold={threshold}: caught {tp}/2 cases, "
f"{fp} false alarms, {fn} missed")
# threshold=0.5: caught 2/2 cases, 1 false alarm, 0 missed
# threshold=0.8: caught 1/2 cases, 0 false alarms, 1 missedSame model. Two thresholds. One catches every stroke but cries wolf once; the other never cries wolf but misses a treatable patient. The "right" answer is a clinical policy, not a coding decision.
Vérification des acquis
1. What does a typical cleared imaging AI triage tool actually output for a given scan?
2. Why is the choice of threshold described as 'the entire business and clinical decision'?
3. For a stroke triage tool detecting large vessel occlusion, why is high sensitivity prioritized over high specificity?
4. Select ALL correct answers about the relationship between sensitivity and specificity when adjusting a model's threshold.
Sélectionnez toutes les réponses correctes.
5. Select ALL correct answers describing what distinguishes imaging AI that improves outcomes from tools that merely 'generate alerts'.
Sélectionnez toutes les réponses correctes.
Here is the part that MBAs underrate and clinicians obsess over. A model with excellent metrics delivers zero value if it does not change what happens next.
The stroke example works because the flag triggers a real workflow: it reorders the worklist and pages a specialist. The value is in time saved to treatment, not in the prediction itself.
Contrast that with a lung nodule tool that emails a report into a system nobody checks until the next morning. Same accuracy, no impact.
There is a subtle risk in tools that work well. When AI is usually right, humans start deferring to it. This is automation bias. A radiologist may under-scrutinize a scan the AI called normal, or over-call one it flagged.
The best deployments preserve human judgment rather than replace it. In practice this often means the AI prioritizes and highlights, but the clinician still makes the diagnosis and owns the decision.
For imaging AI to improve outcomes, every link must hold:
1. Clean ground truth so the model learns the real signal.
2. A validated threshold matched to the clinical use and prevalence.
3. External validation proving it holds up outside the training hospitals.
4. Real workflow integration so a flag changes what happens to the patient.
5. Ongoing monitoring for distribution shift and alert fatigue.
Break any link and the tool becomes expensive shelfware, no matter how good the model looks in a paper.