Pairing-free Group-level Knowledge Distillation for Robust Gastrointestinal Lesion Classification in White-Light Endoscopy

TL;DR

PaGKD tackles the challenge of transferring rich NBI-derived knowledge to WLI-based GI lesion classification without requiring paired NBI-WLI images. It introduces two complementary group-level distillation streams: GKD-Pro, which learns modality-invariant semantic prototypes through a Lesion-Related Query Transformer, and GKD-Den, which enforces dense cross-modal alignment via Semantic Relation-guided Cross-Attention. The framework trains on unpaired WLI and NBI data by operating on grouped samples per class, with losses that encourage semantic consistency and local structural coherence. Across four clinical datasets, PaGKD achieves consistent improvements in AUC compared to state-of-the-art methods, demonstrating strong potential for scalable cross-modal learning in GI endoscopy.

Abstract

White-Light Imaging (WLI) is the standard for endoscopic cancer screening, but Narrow-Band Imaging (NBI) offers superior diagnostic details. A key challenge is transferring knowledge from NBI to enhance WLI-only models, yet existing methods are critically hampered by their reliance on paired NBI-WLI images of the same lesion, a costly and often impractical requirement that leaves vast amounts of clinical data untapped. In this paper, we break this paradigm by introducing PaGKD, a novel Pairing-free Group-level Knowledge Distillation framework that that enables effective cross-modal learning using unpaired WLI and NBI data. Instead of forcing alignment between individual, often semantically mismatched image instances, PaGKD operates at the group level to distill more complete and compatible knowledge across modalities. Central to PaGKD are two complementary modules: (1) Group-level Prototype Distillation (GKD-Pro) distills compact group representations by extracting modality-invariant semantic prototypes via shared lesion-aware queries; (2) Group-level Dense Distillation (GKD-Den) performs dense cross-modal alignment by guiding group-aware attention with activation-derived relation maps. Together, these modules enforce global semantic consistency and local structural coherence without requiring image-level correspondence. Extensive experiments on four clinical datasets demonstrate that PaGKD consistently and significantly outperforms state-of-the-art methods, achieving relative AUC improvements of 3.3%, 1.1%, 2.8%, and 3.2%, respectively, establishing a new direction for cross-modal learning from unpaired data.

Figures (4)

• Figure 1: Overview of the proposed PaGKD. It consists of a pre-trained and frozen NBI classifier, a trainable WLI classifier, and two group-level knowledge distillation modules: Group-level Prototype knowledge distillation (GKD-Pro) and Group-level Dense Knowledge Distillation (GKD-Den). During training, given two unpaired image groups from the same class, GKD-Pro collaborates with GKD-Den to achieve robust and multi-granularity cross-modal knowledge distillation between them.
• Figure 2: t-SNE visualization of the features on PICCOLO. HP (Hyperplasia) $\rightarrow$ AD (Adenoma) $\rightarrow$ ADC (Adenocarcinoma): the severity of the disease increases progressively.
• Figure 3: Ablation on two hyperparameters: $N$ and $s$.
• Figure A1: Receiver operating characteristic (ROC) curves on four datasets: (a) PICCOLO, (b) PolypSet (c) IH-Polyp (d) IH-GC.