Discrete VQ-IHDM: MRI Generation with Vector Quantized Inverse Heat Dissipation Model
Abstract: Accurate and efficient MRI generation is critical in various clinical settings, such as neurology and radiology. The complex data collection procedures, privacy concerns, and lack of medical experts present a bottleneck in the medical imaging data collection and annotation process. In this paper, we adopt a method to unconditionally generate 2D axial brain MRI using a combination of Vector-Quantized image representation and Inverse Heat Dissipation Model (IHDM). We utilize Gaussian Blur as an alternative to order-agnostic masking in the forward process and train a Transformer model to learn the reverse process. This approach allows us to create a single-step sampling algorithm while maintaining high image fidelity. On the ADNI dataset, our model has a FID score of 38.57, a KID score of 0.036, and an ISC score of 1.84.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Mauricio_A_Álvarez1
Submission Number: 1708
Loading