Semantic conflict aware learning improves zero shot chest x ray diagnosis
SCCL-DT: Semantic-Conflict-Aware Contrastive Learning with Dual-Rule Teacher Distillation for Zero-Shot Chest X-Ray Diagnosis
Computer Vision and Pattern Recognition
Summary
Chest X-ray images often have repeated findings across different patients, which can confuse AI training that assumes all different image-text pairs are unrelated. The authors found this causes false negatives during training. They created a new method called SCCL that detects and avoids these mistakes by understanding when different patients actually share similar findings. Their approach combines multiple clever rules into one model that works better at diagnosing chest X-rays without needing extra training on specific diseases.
What this means in practice
- •For healthcare ai developers: Build AI systems for chest X-ray diagnosis that better handle common overlapping findings and improve zero-shot recognition accuracy.
- •For medical imaging platform engineers: Integrate advanced contrastive learning with semantic conflict awareness to improve diagnostic model performance on multi-disease chest X-ray datasets.
Authors
Ting Liu, Haoran Lai, HaobinChen, Ming Liu, S Kevin Zhou
Abstract
Contrastive vision-language pre-training supports zero-shot chest X-ray diagnosis, but standard objectives treat every off-diagonal image-report pair in a batch as a negative. In radiology, recurring findings across patients mean that instance mismatch need not imply semantic incompatibility, creating false-negative supervision that we term a semantic conflict. A training-trace audit shows that pair-observations selected by the online mask used during training are enriched for report-derived finding agreement, although this proxy analysis does not establish clinical equivalence. Fixed thresholds can become miscalibrated as similarity distributions evolve, while soft reweighting leaves residual repulsive gradients. We therefore propose Semantic-Conflict-Aware Contrastive Learning (SCCL), which tracks modality-specific similarity distributions with exponential-moving-average (EMA) statistics, derives adaptive thresholds, and hard-masks flagged candidates before softmax normalization. Inclusive (OR) and consensus (AND) rules define two SCCL teachers, whose bidirectional cross-modal rankings are consolidated through standard distillation into a single deployable student, SCCL-DT. Across five public chest X-ray benchmarks, SCCL-DT achieves a mean zero-shot AUC of $0.850$, compared with $0.837$ for the strongest evaluated baseline in our comparison. Code will be made publicly available soon.