SEEM stumbles on one-pixel vessels: a 0.47M-parameter read-out adapter lifts 3 mm large-vessel Dice to 91.93%
Synopsis
The interactive segmentation model SEEM performs poorly on OCTA retinal vessels out of the box, reaching only 75.83% Dice on large vessels with a frozen encoder, and every candidate mask it can produce is read from the same quarter-resolution feature maps, too coarse for vessels one to three pixels wide. The Read-out Basis Adapter (RBA) adds 32 half-resolution basis maps at the read-out, which each candidate accesses through a zero-initialised projection, for 0.47M trainable parameters. On OCTA-500 3 mm large vessels, averaged over three seeds, it reaches 91.93% Dice and 85.16% IoU, 1.72 and 2.90 points above Conv-LoRA, with gains concentrated on thin vessels and edges and no significant difference on the FAZ.
Interpretation
The adaptation site for interactive retinal vessel segmentation moves from the encoder to the mask read-out: every SEEM candidate mask is a weighted sum of the same quarter-resolution feature maps, and RBA lets candidates draw on a set of half-resolution basis maps built per image. Earlier parameter-efficient methods such as LoRA, Conv-LoRA and AdaptFormer change what the encoder features contain, but not their quarter resolution or the way masks are read from them. On the 50 test images of OCTA-500 3 mm large vessels, RBA beat LoRA rank 16, Conv-LoRA and AdaptFormer on all 50; averaged over three seeds it reached 91.93% Dice against 90.22% for Conv-LoRA, and 85.16% IoU against 82.26%.
RBA has two branches: a read-out branch that builds 32 half-resolution basis maps from the image alone, and an encoder branch, a standard rank-1 LoRA, that supplies the information needed to tell arteries from veins. The read-out projection is initialised to zero, so the model starts exactly as pretrained SEEM and keeps SEEM's candidate generation and selection rule. On the OCTA-500 3 mm and 6 mm subsets and ROSE-O, the gain appears on large vessels, veins and arteries; 3 mm veins reach 90.49% against 88.87%, 6 mm large vessels 88.67% against 85.91%, and ROSE-O vessels 77.55% against 76.23%.
Perspective
Results cover the OCTA-500 3 mm and 6 mm subsets and ROSE-O, with clicks placed by simulation from the ground truth, the model selecting its own candidate at a fixed threshold of 0.5; all training and testing ran on one 8 GB laptop GPU.
Clicks come from error-driven simulation rather than real annotators, and the authors list clicks along vessel boundaries and a study with expert annotators as future work; multi-seed evaluation on the remaining targets and datasets, images from other OCTA devices, and ablations of basis count and resolution are also not yet reported.