On the impact of input resolution on CNN-based gastrointestinal endoscopic image classification

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Gastroenterology
• Computational Pathology

Background:

• Gastric cancer (GC) is a major global health concern.
• Convolutional neural networks (CNNs) show promise in detecting precancerous gastrointestinal (GI) conditions from endoscopic images.
• The impact of endoscopic image resolution on CNN performance for GC detection is not well understood.

Purpose of the Study:

• To investigate the effect of varying image resolutions on CNN performance for intestinal metaplasia (IM) classification.
• To determine the optimal image resolution for training and evaluating CNNs in the context of gastric cancer detection.

Main Methods:

• Utilized two datasets with different resolutions and imaging modalities.
• Trained and evaluated CNN models using input resolutions of 224×224 and 512×512 pixels.
• Assessed model performance using metrics such as F1-scores, considering transfer learning.

Main Results:

• The standard 224×224 pixel resolution is suboptimal for IM detection, even with transfer learning.
• Higher resolutions, specifically 512×512 pixels, consistently yielded superior performance (e.g., InceptionV3 F1-score increased from 91.49% to 94.46%).
• Original image quality and resolution significantly constrain CNN model performance.

Conclusions:

• Maintaining higher original image resolutions from endoscopes is critical for effective CNN training and testing in gastric cancer management.
• Clinicians should prioritize high-quality imaging and consider AI tools that leverage higher resolution data for improved diagnostic accuracy.
• This research informs best practices for AI development and clinical application in endoscopic image analysis for gastric cancer.