Diffusion models for medical image segmentation
Skin lesion segmentation with a diffusion probabilistic model
Diffusion models can produce detailed, high-quality outputs, which makes them appealing for medical image segmentation, where anatomical boundaries must be traced precisely. In this Computer Vision course project, we combined a diffusion probabilistic model with a UNet encoder–decoder to segment skin lesions on the ISIC dataset.
I designed experiments that modified the UNet encoder–decoder architecture and the encoder for the conditioning image, and compared the results against a standard UNet baseline.
Write-up: Report (PDF)