AI-powered hierarchical classification of ampullary neoplasms: a deep learning approach using white-light and...

Endoscopic diagnosis of Ampulla of Vater (AoV) lesions remains challenging owing to complex morphology and limited representative images, particularly for high-risk dysplastic lesions. This study aimed to develop a hierarchical deep learning framework for the stepwise classification of ampullary lesions using white-light (WL) and narrow-band endoscopic images (NBI).

The framework employs three sequential binary classifications: (1) normal vs. abnormal, (2) adenoma vs. cancer, and (3) high-grade dysplasia (HGD) vs. low-grade dysplasia (LGD) within adenomas. Each stage uses EfficientNet-B4 classifiers trained independently on WL and NBI. Predictions are integrated using confidence-based voting. To overcome data scarcity and class imbalance, for HGD and cancer, we used StyleGAN2-ADA to generate synthetic images. The hierarchical model was developed using 4244 endoscopic images from 464 patients collected at Seoul National University Hospital (2693/833/718 for train/validation/test).

The hierarchical model achieved stage-specific accuracies of 95.6% (normal vs. abnormal), 94.4% (adenoma vs. cancer), and 92.7% (LGD vs. HGD), resulting in overall diagnostic accuracy of 92.2%. The model demonstrated excellent sensitivity of 83.3% for HGD and 87.5% for cancer, with specificities exceeding 98%. The confidence-based dual-modality approach (AUROC: 0.921) significantly outperformed single-modality approaches using WL alone (AUROC: 0.866) or NBI alone (AUROC: 0.895), by integrating their complementary diagnostic strengths. Generative adversarial network-based augmentation substantially improved sensitivity for cancer (from 87.5% to 91.7%) and HGD (from 83.3% to 86.5%), while overall accuracy increased from 94.5% to 95.1%.

A hierarchical deep learning approach integrating dual-modality imaging and synthetic data augmentation significantly improves diagnostic performance for ampullary lesions.