| Dec 10, 2024 | My MSc thesis involved the exploration of diffusion models for super-resolution in medical imaging. Specifically Improved, Cold, and ResShift variants, with various backbones including UNet and VIT. These models were tested in both the Latent and Spatial domains, utilizing first-stage models trained in-house, such as VQVAE and VAE. Their performance was compared against established models from hugging face MRI Autoencoder v0.1 and the SDXL-VAE. In the in-house trained VAE, I incorporated a specialized training loss that enabled the model to converge in fewer iterations compared to the traditional combination of GAN loss, KL divergence, and pixel loss. Additionally, I experimented with LoRA to handle data from out-of-distribution data, such as different acceleration factors in MRI images and data variations from FastMRI knee to FastMRI brain and vice versa. I plan to open-source the code and will post extensive analyses on various aspects of this work. Stay tuned for more updates on each topic. |