Radiology report summary - Zero shot

Short Course Q&A ChatGPT Prompt Engineering for Developers
1

I was able to use the demo code to do a zero shot summary of radiology reports. This used to be a hard problem that required lots of annotated data. Here are 3 examples:

Chest: Clear lungs without consolidation
Pleura: No pleural effusion or pneumothorax
Heart: Unremarkable cardiac silhouette
Mediastinum: Unremarkable mediastinal silhouette
Upper Abdomen: NA

Chest: Hyperinflated lungs but clear
Pleura: NA
Heart: Normal size, tortuous aorta
Mediastinum: Normal pulmonary vasculature and hilar contours
Upper Abdomen: NA

Chest: Normal
Pleura: No pleural effusion or pneumothorax
Heart: Normal
Mediastinum: Normal
Upper Abdomen: Nasogastric tube tip in stomach, no other abnormalities detected.

The originals are from RadGraph/data at master · hlk-1135/RadGraph · GitHub
Thank you Andrew for this amazing course!

2

Welcome to the community, A_Stein.
Thank you for sharing this.

3

The important question is whether those analysis are bet-your-life accurate.

4

One of the challenges with radiology report data is that the radiologists mostly do free form dictation. Categorizing reports would allow analysis to look for patterns in a patient population or patterns over time. The accuracy of each individual report classification is therefore not as critical. My plan is to try this out on a larger dataset with over 200K reports and see what comes up :smile: