📢Excited to share our work on #GenerativeAI to explore disparities in large imaging datasets. 🧵(1/5) Read along for a high-level overview, -or- 🎧Listen to the #NotebookLM version. 👇

(2/5) #DiffusionModels, compared to GANs, are very good at capturing data distributions. We trained a model on ~500K pelvic X-rays, conditioning it on demographic info. This allows us to generate highly realistic images representing different patient groups. 🩻

(3/5) We created #counterfactual videos transitioning from one demographic group to another, showing the closest pelvis radiograph. This technique reveals subtle patterns in our dataset without manually reviewing hundreds of thousands of images. 🎥🦴

(4/5) 🔍Key findings: We observed differences in interacetabular distance, osteoarthritis severity, and other features. Notably, a higher degree of osteoarthritis in Black patients points to potential disparities in care access and timing. 🏥 #HealthEquity

(5/5) Read our full article for more details on methodology, findings, and implications for healthcare AI development and #GenerativeAI: Special thanks to @PRouzrokh, first co-author and my partner in crime, @Slowvak, @HillaryGarnerMD, Dr. Doris Wenger, @TauntonM and @CodyWylesMD.

Congratulations @BrdKhsrv and @PRouzrokh !!Solid!!💯💯

👌👌✅✅

Congratulations Bardia. Very well deserved.

Congrats Bardia on this interesting work!👌🏻