Houjun Liu

Diffusion Models for Laproscopic Surgeries

What if we can use diffusion models to generate Laproscopic surgeries to train surgeons?

Problem

Asking dalle to just “generate a Laproscopic surgery” is not going to work. It will give you cartoons.

Approach

  1. text problem formulation: “grasper grasp gallbladder”
  2. encode text into latents
  3. do diffusion with late fusion of latents

Data: Cholec T-45

Weighting

Scoring: Perception Prioritized Weighting + Prioritization for Signal-to-Noise

(Ho et al, 2020)

Text

“[subject] [verb] [object] [surgical phase]”

“grasper grasp gallbladder in preparation”

Model

Elucidated Imagen. Dall-E is very bad; Imagen-class models works better because (why?).

Added Value to Physicians using Generated Images

Train a Classifier

Rendevouz Network: train a discriminator for procedure based on data augmented with generated images; 5% improvement.

Medical Expert Survey

“yo mr doctor man can you spot which one of these are generated?”

45% success rate.