The Catholic University of Korea, Department of Radiology Online Academic Meeting · Medical AI without Coding
A Zoom talk for the department's academic meeting. It reused the 34-slide KSNR autumn symposium talk and added a 10-slide Hands-On section on opening the Claude and Codex apps for the first time, running about 40 minutes including Q&A.
Overview
- Date: Wednesday, 11 November 2026, from 17:30 (about 40 minutes including Q&A)
- Venue: Zoom (online)
- Host: Department of Radiology, The Catholic University of Korea, online academic meeting
- Announced title: Automating medical manuscript writing with AI
- Talk title: Medical AI without coding: A practical introduction
- Format: 44 slides in four sections (Getting Started, Hands-On, For Research, In Practice). The materials folder was shared in the Zoom chat and through a QR code on the last slide
Structure
The talk was built on the 34 slides from the KSNR autumn symposium on 30 October. With more time than a 20-minute slot, I added a new 10-slide Hands-On section on the first things to do after opening the apps.
Section 1, Getting Started, matched the KSNR talk: residents and faculty ask different things of the same tool, seniors who have always defined problems and checked results gain the most as building gets easier, and the shift from chatbot to agent and toward harness engineering.
Section 2: Hands-On
Using screens from the Claude app and Codex (inside the ChatGPT app) that I use daily, I showed how to get started.
- First session. Choosing a working folder from the Code tab in the Claude app, and choosing a project folder in Codex.
- From a CSV to Table 1 and ROC. A 44-second demo recorded during the July course: a synthetic checkup CSV dropped in, with a one-line request.
- The AI's own summary. The AI found and handled all three planted anomalies (a 6.5 kg weight, a fasting glucose of 999, smoking coded "NEVER"). It also left one value based on an unsupported guess, and noted itself that an AUC measured on the same data it was built on is optimistic. Judging those limits takes someone who knows study design.
- Permission modes, model and effort. Start in manual mode, begin large tasks in plan mode, and never use the bypass mode in a folder with patient data.
- Three prompts to save, long chats, Remote Control, and getting unstuck. Ask the AI to question you before starting, ask for evidence, ask for a handover summary; start a new chat or use /compact, keep rules in CLAUDE.md or AGENTS.md; press Esc to stop, and paste error messages as they are.
Sections 3 and 4
Section 3, For Research, was the same as at KSNR: MedSci Skills and its checks, training and externally validating nnU-Net, the npj Digital Medicine and BMC Medical Education papers, the medRxiv preprint, and the reader-study web app with an image quiz. Section 4, In Practice, covered three clinician-built tools (Clinova, Altician, Aperivue RADS), four rules for hospital use, an independent evaluation of an approved aneurysm AI, and how false AI marks affected readers.
Looking ahead
Before the talk I put the slide PDF, a step-by-step course handout, the installation guide, practice data, a one-page cheat sheet, and demo videos in one shared folder. Thanks to the department for the invitation and to the faculty member who arranged the session.