Purpose: Gastrointestinal (GI) endoscopic imaging involves capturing routine anatomical landmarks and suspected lesions during endoscopic procedures for the clinical diagnosis of GI diseases. These images present three key challenges compared to typical scene images: significant class imbalance, a lack of distinctive features, and high similarity between some categories. While existing research has addressed the issue of image quantity imbalance, the challenges posed by indistinct features and inter-category similarity remain unresolved. This study proposes a unified image classification framework designed to tackle all three of these challenges comprehensively.

Methods: We present a novel network architecture, DivGI, which integrates three essential strategies-balanced sampling, fine-grained classification, and multi-label classification-within a single framework. The balanced sampling strategy is implemented via resampling and mix-up techniques, fine-grained classification is enabled through multi-granularity feature learning, and multi-label classification is achieved using hierarchical label joint learning. The performance of our method is validated using three publicly available datasets.

Results: Extensive experimental results demonstrate that DivGI significantly improves classification accuracy compared to existing approaches, with Matthews correlation coefficients (MCC) of 91.31% on the HyperKvasir dataset, 86.72% on the Upper GI dataset, and 82.88% on the GastroVision dataset. These results highlight that DivGI is more effective and efficient compared to existing methods.

Conclusion: The proposed GI classification network, which incorporates multiple strategies, effectively classifies both routine landmark and suspected lesion images, aiming to facilitate better clinical diagnostics in gastrointestinal endoscopy. The code and data are publicly available at https://github.com/howardchina/DivGI.

Keywords: Fine-grained visual recognition; Gastrointestinal endoscopy; Hierarchical label; Image classification; Imbalanced learning.

© 2025. CARS.

目的： 胃肠内镜成像涉及在内镜检查过程中捕捉常规解剖标志和疑似病变，用于临床诊断胃肠道疾病。与典型场景图像相比，这些图像面临三个主要挑战：显著的类别不平衡、缺乏独特特征以及某些类别的高相似性。虽然现有研究已经解决了图像数量失衡的问题，但由非特异性和类别间相似性带来的挑战仍未得到解决。本研究提出了一种统一的图像分类框架，旨在全面应对这三个挑战。

方法： 我们介绍了一个新的网络架构DivGI，该架构在一个单一框架内集成了三种关键策略：平衡采样、细粒度分类和多标签分类。通过重采样和混合技术实现平衡采样策略；通过多粒度特征学习启用细粒度分类；并通过分层标签联合学习实现多标签分类。我们使用三个公开的数据集验证了该方法的性能。

结果： 大量的实验结果显示，DivGI在分类准确性方面显著优于现有方法，在HyperKvasir数据集中马修斯相关系数（MCC）为91.31%，在Upper GI数据集中为86.72%，在GastroVision数据集中为82.88%。这些结果表明，DivGI比现有的方法更有效且高效。

结论： 本文提出的胃肠分类网络通过结合多种策略，有效地对常规标志和疑似病变图像进行分类，旨在促进胃肠道内镜检查中的临床诊断改善。代码和数据可在https://github.com/howardchina/DivGI公开获取。

关键词： 细粒度视觉识别；胃肠内镜；分层标签；图像分类；不平衡学习。