Parameter-Efficient LoRA-GRL Adaptation for Cross-Center Classification of Benign and Malignant Lung Nodules on...

Automated CT lung nodule classification suffers from domain shift across centers. We propose LoRA-GRL, combining Low-Rank Adaptation (LoRA) with adversarial domain harmonization via a Gradient Reversal Layer. Rank-8 LoRA modules were inserted into all attention and feed-forward linear projections (48 sites) of a frozen ViT-S backbone; a domain discriminator encouraged center-invariant features across three centers. Normal vs. Benign and Normal vs. Malignant tasks used 264 patients with patient-level five-fold stratified cross-validation. LoRA-GRL achieved AUCs of 0.9735 and 0.9869, close to full fine-tuning (0.9747 and 0.9863). Differences from strongest baselines fell within overlapping 95% CIs under paired bootstrap, indicating comparable discrimination, not superiority. Efficiency is the main advantage: only 0.789 M trainable parameters, a 96.4% reduction from 21.865 M, and inference latency matched full fine-tuning after adapter merging. Grad-CAM showed nodule-localized predictions but also slice-selection failures. For benign classification, higher AUC than plain LoRA came with higher specificity but lower sensitivity. For malignant classification, sensitivity was 0.9412 (5.9 pp above full fine-tuning) and specificity 0.9697. The small cross-center AUC range (0.0011) reflects internal consistency across participating centers, not unseen-site generalization. LoRA-GRL is a promising parameter-efficient candidate for multi-center lung nodule classification, potentially reducing scanner-specific bias and storage/memory needs, but clinical utility requires external validation and prospective evaluation.