X Longitudinal SSL

blog SSL Longitudinal neighborhood embedding

Order Contrastive Learning¶

Take two images from a patient and predict whether it's in the correct temporal order

Latent Space Alignment¶

Train an autoencoder and learn a unit vector \(\tau\) in the latent space. Force same-subject image pairs to only vary in the \(\tau\) direction ![[Pasted image 20220823110618.png|500]] This is done by adding a cosine loss term which causes the angle between the difference vector of the same-subject image pair and \(\tau\) approach 0. $$ cos(g(I^s; \theta) - g(I^t; \theta), \tau)$$

Longitudinal Neighborhood Embedding¶

MICCAI This approach is generalized to allow for multiple directions by aligning nearest neighbors. Rather than \(\tau\), we simply push the latent space to align with nearby neighbors.

The loss function for LNE is almost the same as for the global alignment, but instead maximizing the cosine between the image pair latent difference and τ, we substitute τ with Δh, a distance-weighted average trajectory of the N nearest embeddings.

This method is appropriate for highly heterogeneous diseases that have different well-defined clusters of trajectories like Alzheimer's.