Local Autoregressive Retinal Vessel Segmentation
A patch-based local autoregressive framework for medical image segmentation — and an honest look at where it breaks.
Autoregressive models generate images token by token. This project asks whether that framing helps for segmentation, by pairing a Vector Quantised Autoencoder with a causal convolutional sequence-to-sequence model over image patches, and training it to produce retinal vessel masks.
In controlled settings the framework reached a 99.88% memorisation rate on the DRIVE dataset. That number is the interesting part, not the impressive part: it is a memorisation result, and the analysis that followed is about the generalisation limits of local autoregressive architectures when training data is scarce — which, for annotated retinal fundus imagery, it always is.
Highlights
- VQ-AE latent space with a causal convolutional seq2seq decoder over patches.
- 99.88% memorisation rate on DRIVE under controlled conditions.
- Analysis of generalisation limits of LAR architectures in the low-data regime.
- Evaluated across DRIVE, CHASE_DB1, STARE and HRF.