Determine why the contrastive objective benefits pen-marking detection

Determine whether the contrastive objective’s improvement on pen-marking detection is caused by the training composition, in which pen marking is abundant, or by properties of latent-space contrastive learning that enable separation of pen marking.

Background

The reconstructed diffusion-based artifact detector shows a reproducible improvement from its auxiliary contrastive term, but the improvement is concentrated on pen marking rather than on tissue folding or air bubbles, which motivated the original contrastive formulation. The authors identify two possible explanations—abundant pen-marking examples in the training data or an intrinsic property of latent-space contrastive learning—but the reported experiment does not distinguish between them.

Resolving this issue would require an experiment that varies artifact composition and separately examines what representations the contrastive objective separates. The paper explicitly leaves the causal explanation unresolved.