Existence of Input-Resolution Constraints Across Alternative Architectures

Determine whether input-resolution operating-range constraints analogous to those identified for a four-stage max-pooling UNet encoder-decoder exist for other retinal vessel segmentation architectures, specifically architectures with larger receptive fields, dilated convolutions, vision transformers, or full-resolution pathways.

Background

The study isolates input resolution as the primary variable using a four-stage UNet and observes that thin-vessel sensitivity depends on the interaction between native image size and encoder pooling depth, revealing an effective operating range of processed widths. The authors note that these findings are specific to the evaluated architecture.

Because many modern segmentation models modify receptive field size or avoid aggressive downsampling (e.g., via dilation, transformer-based designs, or full-resolution pathways), it remains to be established whether similar input-resolution constraints and operating ranges apply beyond the tested UNet configuration.