DivGI: delve into digestive endoscopy image classification

Area of Science:

• Medical Imaging
• Computer Vision
• Gastroenterology

Background:

• Gastrointestinal (GI) endoscopic imaging is crucial for diagnosing GI diseases.
• Existing methods struggle with challenges like class imbalance, indistinct features, and high inter-category similarity in endoscopic images.
• These challenges limit the accuracy of automated image analysis in clinical settings.

Purpose of the Study:

• To propose a unified image classification framework, DivGI, to comprehensively address the challenges in GI endoscopic imaging.
• To improve the accuracy and efficiency of classifying both anatomical landmarks and suspected lesions in endoscopic images.

Main Methods:

• Developed a novel network architecture, DivGI, integrating balanced sampling, fine-grained classification, and multi-label classification.
• Balanced sampling was achieved using resampling and mix-up techniques.
• Fine-grained classification utilized multi-granularity feature learning, and multi-label classification employed hierarchical label joint learning.

Main Results:

• DivGI demonstrated significant improvements in classification accuracy across three public datasets.
• Achieved high Matthews correlation coefficients (MCC): 91.31% on HyperKvasir, 86.72% on Upper GI, and 82.88% on GastroVision.
• Outperformed existing approaches in terms of effectiveness and efficiency.

Conclusions:

• The DivGI network effectively classifies routine and lesion images in gastrointestinal endoscopy.
• The proposed framework facilitates improved clinical diagnostics.
• Code and data are publicly available for further research and application.