Slater-determinant exclusion gating does not improve polyp segmentation: a controlled evaluation with a prespecified analysis plan

Abstract Polyp segmentation networks learn spatially redundant features and often misfire on specular highlights. Determinants give an exact test of redundancy, but layers built on them are seldom compared with controls that isolate the determinant. We built a Slater exclusion gate that scores each feature token by its Schur-complement distance to the span of its neighbours, a score that is zero exactly when the token is linearly dependent on them. On Kvasir-SEG (1,000 images; grouped five-fold cross-validation repeated with three seeds) we compared it with a U-Net, a parameter-matched gate without token interaction and a pairwise-exclusion gate, under an analysis plan fixed before training, and we tested every model without retraining on four external datasets. The gate did not improve per-image Dice over its matched control (0.9006 vs 0.9007; difference − 0.0001, 95% CI − 0.0031 to 0.0029), and it did not reduce false positives inside highlights. Gate diagnostics show why: it attenuated almost every deep token (mean gate 0.08 at stride 32). Adding the same quantity to the loss as a log-determinant penalty lowered inter-token redundancy by 14% with Dice comparable to the U-Net (0.9054 vs 0.9038). Code, splits and verified reimplementations of PraNet and Polyp-PVT are released.