Case Study 2: The Radiologist and the AI — Augmentation in Practice

The Setting

Dr. Elena Vasquez has been a diagnostic radiologist for 14 years. She works at a regional medical center, reading between 50 and 80 imaging studies per day — chest X-rays, CT scans, MRIs. Each study requires her to examine the images systematically, identify any abnormalities, compare them to prior studies when available, and write a report summarizing her findings and recommendations.

It's demanding work that requires years of specialized training, deep pattern recognition, and the ability to maintain focus through hours of screen-based analysis. Radiologists sometimes describe their work as "reading" — a term that captures both the interpretive nature and the sheer volume. In a typical career, a radiologist might read hundreds of thousands of studies.

In 2023, Dr. Vasquez's hospital introduced an AI system designed to assist radiologists with chest X-ray interpretation. The system was developed by a medical AI company, had received regulatory clearance, and had been validated on large datasets. It could analyze a chest X-ray in seconds and flag potential findings — nodules, pneumonia patterns, fractures, enlarged hearts, and other abnormalities — with highlighted regions and confidence scores.

This case study follows the first 18 months of that deployment: what worked, what didn't, and what it reveals about human-AI augmentation in practice.

Before the AI: A Day in Dr. Vasquez's Work

To understand the AI's impact, we first need to understand the workflow it entered.

Before the AI, Dr. Vasquez's typical day looked something like this:

• 7:00 a.m.: Arrive, review overnight urgent studies that were preliminary-read by overnight residents.
• 7:30 a.m. – 12:00 p.m.: Read studies from a worklist. Each chest X-ray takes 2–5 minutes to interpret; CT scans take 10–30 minutes depending on complexity.
• 12:00 – 12:30 p.m.: Lunch (often at her workstation).
• 12:30 – 5:00 p.m.: Continue reading. Periodic interruptions for consultations with referring physicians who call to discuss findings.
• Throughout the day: Document findings by dictating reports using speech recognition software.

The bottleneck was volume. The hospital's imaging volume had grown by roughly 30% over the preceding five years, while the radiology department hadn't grown proportionally. Dr. Vasquez and her colleagues were reading more studies per day than they had five years ago, with less time per study.

Fatigue was a real concern. Research has shown that radiologist accuracy declines over the course of a long reading session. The most dangerous errors often happen late in the day, when attention flags and the studies blur together. "Satisfaction of search" — the tendency to stop looking after finding one abnormality, missing a second one — is a well-documented cognitive bias in radiology.

The AI Arrives: Implementation

The hospital implemented the AI system with a deliberate augmentation-first approach. The system was explicitly designed not to replace radiologists — it was designed to serve as a "second reader" that could:

1. Pre-screen studies. Flag chest X-rays with likely abnormalities so they could be prioritized in the worklist (rather than read in chronological order).
2. Highlight regions of interest. Overlay markers on images where the AI detected potential findings, with associated confidence scores.
3. Catch near-misses. After the radiologist completed their read, the system could flag additional findings that the radiologist might have overlooked.

Crucially, the system was not given the authority to make final diagnoses. Every study still required a radiologist's interpretation and sign-off. The AI was an assistant, not a decision-maker.

What Worked

Prioritization

The most immediately useful feature was the prioritization of urgent findings. Before the AI, a critical finding — a collapsed lung, a large mass, a widened aorta suggesting potential rupture — sat in the worklist in chronological order. If it was submitted at 10:00 a.m. and the radiologist was working through studies from 8:00 a.m., it might not be read until afternoon.

The AI flagged high-priority studies within seconds of image acquisition, moving them to the top of the worklist. This reduced the time-to-diagnosis for critical findings significantly. In the first year, the radiology department documented multiple cases where the AI's prioritization likely improved patient outcomes by accelerating the identification of time-sensitive findings.

Dr. Vasquez describes this as the clearest win: "The AI doesn't read the image better than I do. But it reads it faster. For a patient with a tension pneumothorax, those minutes matter."

Catching Subtle Findings

The AI also proved useful for catching findings that were genuinely difficult to see. Small pulmonary nodules — tiny spots in the lung that might indicate early cancer — are notoriously easy to miss on chest X-rays, even for experienced radiologists. The AI's highlight function drew attention to regions that Dr. Vasquez might otherwise have dismissed or overlooked.

"I've had cases where the AI flagged something I initially disagreed with," she says. "I looked more carefully, and it was right. A small nodule I would have missed. That's the AI doing exactly what it's supposed to do — making me better at my job."

Reducing Cognitive Load

After 18 months, Dr. Vasquez noticed a subtler benefit: reduced cognitive load for routine studies. When the AI scanned a chest X-ray and returned "no significant findings — confidence 0.97," she still reviewed the image herself. But knowing that a sophisticated pattern-recognition system had already screened it allowed her to approach the review differently — more as a verification task than an exhaustive search. This felt less draining, particularly late in the day.

What Didn't Work

Over-reliance Risk (Automation Bias)

Three months into the deployment, the department noticed a concerning pattern. Some radiologists — particularly the more junior ones — were spending less time on studies that the AI flagged as normal. They were trusting the AI's assessment and performing a cursory review rather than a thorough one.

This is a well-documented phenomenon called automation bias: the tendency to over-trust automated systems, particularly when they provide quantified confidence scores. An AI that says "normal — 97% confidence" is psychologically powerful. It creates a sense of safety that can be false.

The department conducted an internal audit and found that the AI's false-negative rate — the rate at which it failed to flag genuine abnormalities — was low but non-zero. In the first year, the AI missed findings in approximately 3–4% of studies that had true positives. Most of these were subtle findings that were also difficult for human readers. But in a few cases, the missed finding was clinically significant.

The department responded by implementing a policy: radiologists were required to complete their own read before reviewing the AI's output, to prevent the AI's assessment from anchoring their interpretation. This reduced the over-reliance pattern but also reduced the cognitive load benefit — a genuine trade-off.

Alert Fatigue

The AI generated a lot of alerts. In its eagerness to avoid missing anything (high sensitivity), it flagged many findings that turned out to be insignificant — old fractures that had healed, calcified granulomas from previous infections, normal anatomical variants that happened to look unusual to the algorithm.

After a few months, Dr. Vasquez noticed herself dismissing the AI's flags more quickly. "It's the boy who cried wolf," she says. "When 40% of the flags are things I already know about or that aren't clinically significant, I start paying less attention to all of them. That's dangerous, because the important flags get lost in the noise."

This alert fatigue is a common problem in clinical decision support systems. The system's designers face a fundamental tension: set the sensitivity too high, and you get too many false alarms; set it too low, and you miss real findings. Finding the right threshold requires ongoing calibration based on the specific clinical context and feedback from the radiologists using the system.

The "Deskilling" Concern

A more philosophical concern emerged over time. Several of the senior radiologists, including Dr. Vasquez, worried about the long-term effects on resident training. Radiology residents traditionally develop their skills by reading thousands of studies and building an internal library of patterns. If AI pre-screens studies and highlights regions of interest, will residents develop the same depth of independent pattern recognition?

"I learned by struggling," Dr. Vasquez says. "I read a chest X-ray, missed the nodule, got corrected by an attending, and never missed that pattern again. If the AI always highlights the nodule for you, do you ever learn to find it yourself? And if the AI goes down one day — a system failure, a cybersecurity event — will you be able to read without it?"

This concern doesn't have a clear resolution yet. It represents a genuine tension between short-term performance (AI makes radiologists better right now) and long-term capability (AI might make future radiologists less skilled as independent practitioners).

The Bigger Picture: Augmentation as a Model

Dr. Vasquez's experience illustrates several principles that apply far beyond radiology:

1. Augmentation Requires Deliberate Design

The hospital's success — such as it was — came from deliberately designing the AI as an assistant rather than a replacement. The system was integrated into existing workflows, not imposed as a substitute for human judgment. The radiologists retained final authority. This matters: the same AI technology, deployed differently (as an automated screening service that only escalated to humans for complex cases), would have created a fundamentally different — and potentially more dangerous — system.

2. The Interface Matters as Much as the Algorithm

How the AI's output is presented to the human affects how the human uses it. A confidence score of "97% normal" has a powerful anchoring effect. A simple "no AI flags — verify independently" might be more effective at preventing over-reliance. The user interface design of human-AI systems is an under-appreciated factor in their real-world performance.

3. Augmentation Changes the Job Even Without Eliminating It

Dr. Vasquez still reads X-rays. She still writes reports. She still consults with referring physicians. But the texture of her work has changed. She spends less time on routine screening and more time on complex cases. She interacts with an AI system throughout her day. She thinks about automation bias and alert fatigue. Her professional identity is shifting from "expert image reader" to "expert image reader who collaborates with AI."

4. The Benefits Are Real but Come with Trade-offs

Faster prioritization saves lives. Catching missed nodules improves outcomes. Reduced cognitive load helps radiologists maintain performance over long shifts. These are genuine benefits. But they come packaged with automation bias, alert fatigue, deskilling concerns, and the ongoing cost of managing the human-AI relationship. Augmentation is not a free lunch.

Discussion Questions

1. Dr. Vasquez's hospital required radiologists to complete their own interpretation before reviewing the AI's output. What are the advantages and disadvantages of this policy? Can you think of a better approach?

2. The "deskilling" concern applies beyond radiology. Think of another profession where AI augmentation might improve current performance while potentially reducing the development of human expertise. How would you address this tension?

3. The AI's alert fatigue problem reflects a fundamental trade-off between sensitivity (catching everything) and specificity (avoiding false alarms). How should this trade-off be resolved, and who should make that decision — the AI developer, the hospital, the individual radiologist, or someone else?

4. This case study presents augmentation as largely positive. What would a negative augmentation story look like — a case where AI augmentation made a professional's work worse even without replacing them?

5. How does this case compare to Case Study 1 (the warehouse)? Both involve AI changing how humans do their work. What's different about the experience, and why?

Connection to Your AI Audit Report

Consider whether the AI system you're auditing could be deployed in an augmentation model (like Dr. Vasquez's experience) or a replacement model. Ask yourself: - Who would the AI assist, and what would their experience be like? - What safeguards would prevent over-reliance (automation bias)? - How would the human-AI interface be designed to support good decisions rather than anchoring bad ones? - What skills might degrade over time if humans rely on the AI?