AI doesn’t replace radiologists. Radiologists using AI will replace those who don’t.
A conversation with Samir Abboud, Chief of Emergency Radiology at Northwestern Memorial Hospital.
Radiologists using AI will replace those who don’t—but only when AI is designed around real clinical workflows, quality safeguards, and human judgment.
That’s the real lesson from what happened at Northwestern Medicine.
Not that AI can boost productivity.
Not that generative models can draft radiology reports.
But that clinical AI only works when it removes the right kind of work—and leaves responsibility exactly where it belongs.
At Northwestern Memorial Hospital, an internally built generative AI tool helped radiologists work meaningfully faster in real clinical practice—without a measurable drop in report quality. In some cases, much faster. And it did so without automating diagnosis, bypassing clinicians, or forcing new workflows.
This piece breaks down what actually changed, why adoption stuck, and what this case reveals about why most clinical AI struggles to move beyond pilots.
Executive Summary
Dr. Samir Abboud, Chief of Emergency Radiology at Northwestern, co-developed ARIES (Automated Radiology Interpretation and Evaluation System), an in-house generative AI system.
In a June 2025 JAMA study, radiologists using ARIES achieved average efficiency gains of ~15%, with some approaching ~40%, without degrading report quality, as assessed by blinded peer review.
The gains came from eliminating clerical and cognitive overhead—not from reading images faster.
Adoption scaled through clinician champions, personalization, and transparency—not mandates.
The core lesson: AI doesn’t replace radiologists. Radiologists using AI will replace those who don’t—when AI is built around real workflows, quality safeguards, and human judgment.
🎧 Watch the full episode:
Takeaway 1: The Bottleneck Was Never Diagnostic Skill
One of the most important clarifications Dr. Abboud made was also the simplest.
In many cases, radiologists know what’s going on with a patient very quickly after reviewing the images. The time sink comes afterward—dictating the report, fixing grammar, editing templates, and refining language that doesn’t change the diagnosis but still demands attention.
ARIES was designed to target that exact gap.
The system ingests the current study, prior images, and basic metadata, then generates a full narrative draft report—what Abboud describes as “pixels to paragraphs.” That draft appears directly inside the existing dictation workflow.
In practice:
80–85% of X-ray reports required no edits
~10% needed minor adjustments
~5% were faster to complete manually
The efficiency gains didn’t come from speeding up interpretation. They came from removing work that added cognitive load without adding clinical value.
Implication: Many healthcare AI tools fail because they try to augment the most complex part of clinical work. ARIES succeeded by eliminating the invisible, draining parts instead.
Takeaway 2: Speed Was Irrelevant Without Proving Quality
Northwestern didn’t treat productivity gains as inherently positive. They tested whether speed came at a cost.
The JAMA study tracked:
Time from opening a study to signing a report
Differences with and without AI assistance
Independent review of 800 reports, with reviewers blinded to AI use
Reviewers assessed missed findings, clarity, grammar, and interpretability. The result: no evidence of quality degradation.
Abboud was explicit about why this mattered. Anyone can go faster and do a worse job. The challenge was maintaining the standard patients expect while reducing friction for clinicians.
Implication: In healthcare, AI that cannot demonstrate quality preservation under real conditions will not scale—no matter how impressive the model looks.
Takeaway 3: Adoption Was Won Peer-to-Peer, Not Top-Down
ARIES didn’t spread because leadership mandated it.
Early adoption happened almost accidentally. Radiologists noticed AI-generated text appearing in their workflow and wanted access. That curiosity created a natural group of early adopters.
Northwestern formalized this by identifying clinician “champions” across subspecialties and community hospitals. These users received early access and hands-on support, then shared results informally with peers.
A shared dashboard showed anonymized efficiency gains over time—allowing radiologists to see learning curves improve month by month without shaming or competition.
Over roughly six months, usage grew from a single user to about 60 active radiologists across 11 hospitals.
Implication: In clinical environments, trust flows laterally. AI adoption follows peer credibility, not org charts.
Takeaway 4: The Holdouts Exposed the Real Friction
Some radiologists resisted ARIES—not because it made diagnostic errors, but because it didn’t sound like them.
They spent time editing phrasing rather than substance. The friction was stylistic, not clinical.
Instead of dismissing those complaints, the team treated them as signal. Engineers added a layer that allowed reports to reflect each radiologist’s preferred language. Adoption increased. Efficiency improved. Friction dropped.
Abboud described these “picky” users as essential contributors to making the system usable at scale.
Implication: Resistance often reveals where technology clashes with professional identity. Fixing that alignment can unlock adoption faster than adding new features.
Takeaway 5: Context Matters More Than Detection
Most imaging AI tools focus on detecting a single condition extremely well. The unintended consequence is alert fatigue.
ARIES takes a different approach. Because it generates full reports, it can reason over clinical context.
Abboud gave the example of pneumothorax. Many ICU patients have expected or already-treated pneumothoraces. Traditional systems still flag them. ARIES filters alerts to surface only new, enlarging, or unexpected cases.
Instead of dozens of alerts per day, clinicians saw one or two per week—and knew they mattered.
Implication: Clinical AI that understands when something is important is far more valuable than AI that simply detects that something exists.
Why This Matters
This isn’t really a radiology story.
It’s a case study in why most healthcare AI struggles to escape pilots—and why a small number of tools quietly become indispensable. The difference isn’t accuracy claims or autonomy. It’s respect for workflow, accountability, and human judgment.
Dr. Abboud’s view of the future is pragmatic: AI will not replace radiologists. But radiologists who use AI—critically, thoughtfully, and with proper safeguards—will outpace those who don’t.
For HealthTech founders, operators, and investors, the lesson is sobering and actionable at the same time: the hardest part of clinical AI isn’t building the model. It’s designing for the humans who remain responsible for every decision.
Thanks for reading. Stay healthy,
Rodrigo Hütt