Cross-Window Distillation Boosts Lung CT AUC to 0.996

A team of researchers led by Bo Peng developed a cross-window knowledge distillation framework that significantly improves pulmonary CT analysis. Traditional deep learning methods fuse multi-window CT representations only at later stages, missing cross-density interactions. The new approach uses a teacher-student setup: a teacher trained on the most informative window imparts latent clinical priors to student encoders for other windows. This distills the teacher's knowledge, allowing students to internalize pathological signatures invisible to standard supervised training.

Evaluated on three large cohorts (COPD-CT-DF with 719 samples, RSNA PE with 1,433, and an in-house CTEPD dataset of 161), the method yielded dramatic AUC gains. Per-window AUC on COPD-CT-DF jumped 10.1–16.5 percentage points (from 0.75–0.81 to 0.90–0.94, all P<0.001), with ensemble AUC reaching 0.9960. Similar improvements were seen on RSNA PE (0.80–0.83 to 0.90–0.92) and CTEPD (0.7481 vs. 0.6264). The framework offers a generalizable solution for multi-window pulmonary CT analysis, unlocking latent pathological signatures that supervised methods miss.