[[
wikihub
]]
Search
⌘K
Explore
People
For Agents
Sign in
Explore
People
For Agents
Sign in
@jemoka / Jemoka Knowledge Base / raw/concept/kbhclinical_skin_disease_imaging.md
Suggest edit
Cancel
Submit suggestion
Title
Name
Note
--- title: "Clinical Skin Disease Image Generation" source: https://www.jemoka.com/posts/kbhclinical_skin_disease_imaging/ --- key question: are there bias and changes in young children + changes in skin color to generate more samples of skin disease. previous work: DermGAN (Ghorbani 2020), this is not pediatric and also a bit deterministic. key problems data is scarce data is not available and lack of data sharing. data is sensitive especially children. pediatric specificity we want to generate children’s skin disease samples, which as vastly out of sample. The work is therefore trained on only 1000-2000ish samples. modeling latent diffusion model (LDF) ControlNet (Zhang 2023)—allows specific conditioning of the generation by exogenous rules data gym cleaning get rid of face crop for specific body part ensure anonymity, etc. patch extraction start in the upper left corner, ensure that the resulting patch has at least n% of diseases available further, create a mask for where the disease should live modeling train a diffusion model yay results the diffusion model by itself does nothing combined with ControlNet, life is much better and achieve higher than SOTA