Lesion Detection in CT with Frozen Self-Distilled Features: SALT, a Spatially Adaptive Label-Guided Temperature
The paper introduces SALT, a spatially adaptive self-distillation method for CT lesion detection. During pretraining, weak box-derived labels identify a compact region on the encoder’s patch grid. Inside that region, SALT sharpens the teacher temperature and increases the masked-patch loss weight, while leaving the masking policy, centering statistics, architecture, and schedule unchanged. At downstream use, the encoder is frozen and operates without labels or conditioning; only a lightweight multi-depth CenterNet-style 3D detection head is trained. The study proposes evaluations across four CT cohorts, lesion-size strata, box quality, and a detector-free follow-up re-identification probe.
Why it's worth reading
SALT is worth reading now because it changes the self-distillation target itself, rather than merely sampling lesion-centered views, potentially improving representations for small lesions. The claimed gains still require verification from the full quantitative results.