SCCL-DT: Semantic-Conflict-Aware Contrastive Learning with Dual-Rule Teacher Distillation for Zero-Shot Chest X-Ray Diagnosis
The SCCL framework fixes false negatives in zero-shot chest X-ray diagnosis by adaptively masking similar image-report pairs that represent semantic conflicts. It tracks similarity distributions using exponential moving averages to set dynamic thresholds, preventing fixed thresholds from becoming miscalibrated. The resulting model, SCCL-DT, achieves a mean zero-shot AUC of 0.850 across five public benchmarks.