Basis

The architecture we are using is outlined in this paper. Originally, we were planning on using a GAN-based model. But, as outlined in the paper, diffusion-based models have proven to be more clinically relevant.

Brownian Bridge Diffusion Model

Diffusion consists of two processes, the forward and backwards processes. The forward process consists of applying noise and gradually destroying information of the image, and the backwards process is the removal of noise from an image until our synthetic image is all that remains. So when we want to perform an inference, we give the model pure noise and it does denoising until an image is present. This video does a fantastic job of understanding this process. Specifically, we use a conditional diffusion model, meaning that we give it a starting point, the model applies noise, then denoises; in contrast to just denoising from pure noise. Information about denoising is traditionally learned using a U-net backbone.

The Browian Bridge in this context refers to the noising schedule, instead of the end point of the noising step being complete and pure noise, it still has some information from the input. The Brownian Bridge keeps training and inferences stable. Brownian Bridge excels in situations where there are known endpoints, in our case, a CT and and MRI. Additionally, we get the benefit of better signal preservation, since information is never completely destroyed, we get better preservation of brain structures.

Style Key Conditioning and Inter-Slice Tradjectory Alignment Sampling

In addition to the U-net backbone, the paper also applies cross-attention, informing the model of scanner-specific style consistencies. The style cross-attention is called Style Key Conditioning(SKC) in the paper. In addition to this, the paper also proposes Inter-Slice Trajectory Alignment (ISTA) sampling. This technique consists of using co-prediction, using an inference with the model and taking the average of surrounding slices so that inter-slice structures are consistent. Then it performs a correction by calling another inference, since this is a sampling technique, it only applies to inference-time, there are no training trade offs that need to be made here.