Persistent Cross-Institutional Domain Shift in Narrow-Band Imaging Classification of Colorectal Lesions and the Utility of Site-Specific Fine-Tuning

Background/Objectives: Cross-institutional transfer of endoscopic deep learning models may be limited by acquisition and preprocessing differences. We reassessed persistent domain shift after correcting Facility B input linkage and validation design. Methods: A preflight audit corrected two Facility B image–mask linkage omissions without changing the 402 source lesions or 469 still images. Criterion B was a qualitatively selected, Facility-B-informed HSV patch-selection adjustment; the same Tenengrad focus criterion was used at both facilities. Class-wise MMD2 was calculated using all three Facility A fold-specific feature extractors. Each Facility A source model was evaluated on the same corrected criterion B Facility B cohort. Facility B models were initialized from ImageNet and fully fine-tuned for four prespecified epochs under lesion-grouped three-fold cross-validation, with pooled out-of-fold patch-level evaluation. Historical Facility A internal results were descriptive only. Results: Criterion B retained 142,878 Facility B patches. Residual MMD2 was greatest for SM1 (0.185) and SM2 (0.147). Cross-site accuracy, macro-F1, and balanced accuracy were 0.711, 0.397, and 0.422; site-specific values were 0.770, 0.472, and 0.473. Recall improved for BG, LG, and SM1 but decreased for HG and SM2. Conclusions: Site-specific development partially improved aggregate local performance, but persistent domain shift and class-specific limitations remained; universal cross-institutional generalizability was not established.