The AI tools that touch Mr. Diaz's scan, how they share the work with radiologists, AI in the rest of medicine, and generative AI in the clinic
Module 1 explained how AI works. Module 2 explained how radiology works. This module puts them together and asks what AI changes for patients and for you. We start by looking at the AI tools in use today, first in radiology, then across medicine. About 15-20 minutes.
By the end of this lesson you can:
In Lesson 2.2 you followed Mr. Diaz's CT pulmonary angiogram (CTPA) from the order to the follow-up. Now imagine the same night at a hospital that uses AI. None of these tools replaces a person. Each one does one narrow job at one step.
| Step (Lesson 2.2) | AI tool | What it does for Mr. Diaz |
|---|---|---|
| 1 · Order | Decision support | Checks the order against appropriateness guidelines and suggests CTPA for suspected PE. |
| 3 · Scan | Image reconstruction | Cleans up noise so a lower radiation dose still gives a clear picture. |
| 5 · Worklist | PE triage | Spots a likely clot within minutes and moves his scan to the top of the radiologist's list. |
| 6 · Read | Nodule detection | Marks the 8 mm lung nodule and measures it. |
| 7 · Report | Report drafting | Suggests impression wording from the radiologist's findings, and a plain-language version for the patient portal. |
| 9 · Follow-up | Follow-up tracker | Reads the recommendation and reminds the team if the 6-12 month CT is not booked. |
A radiology worklist (screenshot or mock-up, no patient details) with one CTPA flagged by a PE triage AI and moved to the top.
An axial chest CT slice with a nodule-detection AI box and measurement around a small lung nodule.
Most radiology AI does one of four jobs: detect (is something there?), triage (which scans are urgent?), measure (how big, how much?) or draft (put words to it). Each tool on the list was trained and tested separately. A tool that is good at clots knows nothing about nodules.
At the reading step, the same AI tool can be set up in different ways. What changes is when the radiologist sees the AI output, and who has the last word.
For each card, guess when the radiologist sees the AI's output, then flip it.
The triage tool misses Mr. Diaz's clot. His scan stays in the normal queue. Does he still get a correct diagnosis? (Yes, but later: a radiologist still reads it. A miss by a triage tool costs time, not the diagnosis.)
Radiology has the most AI tools of any specialty, but it is not the only one. The ideas from Module 1 apply everywhere: a model learned from examples, it has a sensitivity and specificity, and it can fail on patients unlike the ones it learned from.
Any specialty that looks at pictures
ECGs and wearables
Numbers in, risk scores out
Generative AI and large language models
Most of the tools above are narrow: one task, one kind of input, one kind of answer. Generative AI is different. A large language model (LLM) learned from huge amounts of text to predict the next word (Lesson 1.1), and it can write almost anything: a note, a summary, a letter, an answer to a question.
That makes it useful and risky at the same time:
An LLM wrote this plain-language summary of Mr. Diaz's CTPA report (Lesson 2.4) for the patient portal. Three of its six sentences are wrong. Click each sentence to check it against the real report: PE in the right lower lobe, no strain on the heart, 8 mm nodule, follow-up CT in 6-12 months.
Decide in your head whether it matches the report before you click.
A segmental pulmonary embolism in the right lower lobe, put into plain words.
The report says "no CT evidence of right heart strain." The model may have picked up the topic of heart strain and got it backwards. This would frighten the patient and could mislead a reader who trusts the summary.
The 8 mm right upper lobe nodule, in plain words.
The report never calls the nodule a scar and recommends a follow-up CT in 6-12 months. This is the most dangerous error here: it sounds kind, and it could stop Mr. Diaz from getting the scan that would catch an early cancer.
Not in the radiology report itself, but true to the situation and phrased so the care team decides. A good summary points to the next conversation rather than making promises.
This guideline and paper do not exist. LLMs can produce citations that look real, with authors and years. Look up every reference before you trust or repeat it.
Four rules for using generative AI:
In the United States, AI that diagnoses or guides treatment is regulated as a medical device by the Food and Drug Administration (FDA). The FDA keeps a public list of AI-enabled devices. By mid-2026 it held more than 1,600 devices, and about three in four were for radiology.
The AI2Practice AI Tool Tracker summarizes that list: how many tools there are, which specialties and tasks they cover, and how the numbers have grown each year. Most are narrow tools like the ones in Mr. Diaz's table. Note what the list does not tell you: whether a tool helps patients, or whether it works at your hospital. That is the topic of Lesson 3.2.
Open the tracker and find the specialty with the second-most tools. Then pick one radiology task, such as stroke or mammography, and guess how many authorized tools do it.
Answer each question to continue. Any answer is fine; read the explanation.
A PE tool moves likely urgent scans to the top of the worklist, and the radiologist still reads every scan. Which reading design is this?
Triage changes the order of reading, not who reads. If it misses a clot, the scan is read later rather than never.
An LLM summary of a report includes a reference with authors and a year. What should you do?
LLMs can invent references that look real. They can still be useful drafts, as long as you check facts and citations against the source.
Which of these is safe to paste into a free public chatbot?
A general question with no patient details is fine (and you should still check the answer). Reports, names, dates and screenshots are protected health information and only go into approved tools.