Holistic White-light Polyp Classification via Alignment-free Dense Distillation of Auxiliary Optical Chromoendoscopy

Abstract: White Light Imaging (WLI) and Narrow Band Imaging (NBI) are the two main colonoscopic modalities for polyp classification. While NBI, as optical chromoendoscopy, offers valuable vascular details, WLI remains the most common and often the only available modality in resource-limited settings. However, WLI-based methods typically underperform, limiting their clinical applicability. Existing approaches transfer knowledge from NBI to WLI through global feature alignment but often rely on cropped lesion regions, which are susceptible to detection errors and neglect contextual and subtle diagnostic cues. To address this, this paper proposes a novel holistic classification framework that leverages full-image diagnosis without requiring polyp localization. The key innovation lies in the Alignment-free Dense Distillation (ADD) module, which enables fine-grained cross-domain knowledge distillation regardless of misalignment between WLI and NBI images. Without resorting to explicit image alignment, ADD learns pixel-wise cross-domain affinities to establish correspondences between feature maps, guiding the distillation along the most relevant pixel connections. To further enhance distillation reliability, ADD incorporates Class Activation Mapping (CAM) to filter cross-domain affinities, ensuring the distillation path connects only those semantically consistent regions with equal contributions to polyp diagnosis. Extensive results on public and in-house datasets show that our method achieves state-of-the-art performance, relatively outperforming the other approaches by at least 2.5% and 16.2% in AUC, respectively. Code is available at: https://github.com/Huster-Hq/ADD.

Holistic White-light Polyp Classification via Alignment-free Dense Distillation of Auxiliary Optical Chromoendoscopy

The paper "Holistic White-light Polyp Classification via Alignment-free Dense Distillation of Auxiliary Optical Chromoendoscopy" addresses the pressing need for improved diagnostic accuracy in colorectal polyp classification, a critical element in colorectal cancer screening. Colorectal cancer remains a leading cause of cancer-related deaths, and effective early detection relies on accurate polyp classification during colonoscopy. Typically, colonoscopy utilizes White Light Imaging (WLI) as the standard modality, complemented by Narrow Band Imaging (NBI), which provides enhanced vascular detail through optical chromoendoscopy. Despite the advantages of NBI, WLI is often the only available modality in resource-limited settings, necessitating advancements in WLI-based approaches to bridge the accuracy gap with NBI.

The research presented introduces a novel framework that enhances WLI classification by leveraging the rich feature data from NBI, without necessitating direct image alignment or polyp localization. This is achieved through the innovative Alignment-free Dense Distillation (ADD) module, which performs fine-grained cross-domain knowledge distillation. This module circumvents alignment issues prevalent between WLI and NBI images, establishing reliable pixel-wise associations across feature maps, thus facilitating the transfer of diagnostic knowledge.

Key to the ADD module's function is the establishment of pixel-wise affinities that guide distillation across misaligned images. This approach ensures the distillation process connects semantically relevant regions, employing Class Activation Mapping (CAM) to refine these affinities and maintain diagnostic consistency.

Experimental results from trials on both public and proprietary datasets demonstrate significant improvements in WLI classification accuracy and reliability, with relative boosts in AUC by at least 2.5% on standard test cases and up to 16.2% in controlled internal datasets. These results are noteworthy, indicating a substantial leap forward in the accuracy of WLI-based diagnostic methods.

The paper also compares the proposed methodology with existing approaches, including global alignment techniques and cropped lesion analysis methods, demonstrating superior performance by effectively overcoming shortcomings associated with detection errors and loss of context. This contextual strength in diagnosis as facilitated by holistic image analysis, as opposed to lesion-centric models, underscores the importance of comprehensive data utilization in medical imaging models.

The implications of this research are multifaceted, offering practical enhancements in clinical diagnostic settings, particularly where equipment and resource constraints limit the availability of advanced imaging modalities. Additionally, on a theoretical level, the paper contributes to discussions on cross-domain learning and knowledge distillation frameworks, potentially influencing future developments in AI for medical imaging.

Future work may explore scaling the framework for various imaging conditions and modalities beyond colonoscopy, further exploring the domain adaptation potential to facilitate accurate real-time diagnostics in diverse medical contexts. This approach is thus poised to influence both current practices and future innovations in medical imaging technology and AI applications in healthcare.