kb autocommit

content/posts/KBhaaai2024_index.md

@@ -4,7 +4,14 @@ author = ["Houjun Liu"]
 draft = false
 +++
 
-| Locale   | Speaker           | Topic + Link                                                              |
-|----------|-------------------|---------------------------------------------------------------------------|
-| W3PHI-AI | Elizabeth Broycki | [AI Healthcare Safety]({{< relref "KBhai_healthcare_safety.md" >}})       |
-| W3PHI-AI | Jeff Clark        | [Patient Risk Prediction]({{< relref "KBhpatient_risk_prediction.md" >}}) |
+| Locale   | Speaker             | Topic + Link                                                                                         |
+|----------|---------------------|------------------------------------------------------------------------------------------------------|
+| W3PHI-AI | Elizabeth Broycki   | [AI Healthcare Safety]({{< relref "KBhai_healthcare_safety.md" >}})                                  |
+| W3PHI-AI | Jeff Clark          | [Patient Risk Prediction]({{< relref "KBhpatient_risk_prediction.md" >}})                            |
+| W3PHI-AI | Yasmine and Emily!  | [Abulance Trajectories]({{< relref "KBhabulance_trajectories.md" >}})                                |
+| W3PHI-AI | Simeon Allmendinger | [Diffusion Laproscopic Surgeries]({{< relref "KBhdiffusion_models_for_laproscopic_surgeries.md" >}}) |
+
+
+## Talk Contact {#talk-contact}
+
+[AAAI Talk Contacts]({{< relref "KBhaaai_talk_contacts.md" >}})

content/posts/KBhdiffusion_models_for_laproscopic_surgeries.md

@@ -0,0 +1,55 @@
++++
+title = "Diffusion Models for Laproscopic Surgeries"
+author = ["Houjun Liu"]
+draft = false
++++
+
+What if we can use diffusion models to generate Laproscopic surgeries to train surgeons?
+
+
+## Problem {#problem}
+
+Asking dalle to just "generate a Laproscopic surgery" is not going to work. It will give you cartoons.
+
+
+## Approach {#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 {#weighting}
+
+Scoring: Perception Prioritized Weighting + Prioritization for Signal-to-Noise
+
+(Ho et al, 2020)
+
+
+### Text {#text}
+
+"[subject] [verb] [object] [surgical phase]"
+
+"grasper grasp gallbladder in preparation"
+
+
+### Model {#model}
+
+Elucidated Imagen. Dall-E is very bad; Imagen-class models works better because (why?).
+
+
+## Added Value to Physicians using Generated Images {#added-value-to-physicians-using-generated-images}
+
+
+### Train a Classifier {#train-a-classifier}
+
+Rendevouz Network: train a discriminator for procedure based on data augmented with generated images; 5% improvement.
+
+
+### Medical Expert Survey {#medical-expert-survey}
+
+"yo mr doctor man can you spot which one of these are generated?"
+
+45% success rate.