Binary polyp-size classification based on deep-learned spatial information.

Area of Science:

• Medical imaging
• Gastroenterology
• Artificial intelligence

Background:

• Accurate polyp size estimation is crucial for colon cancer screening and determining surveillance intervals.
• Endoscopists' subjective size estimations can be inaccurate, potentially leading to misdiagnosis.
• Distinguishing between smaller (1-9 mm) and larger ([Formula: see text] mm) polyps is clinically significant.

Purpose of the Study:

• To develop an automated method for binary polyp-size classification.
• To differentiate between polyps measuring 1-9 mm and those measuring [Formula: see text] mm.
• To improve the accuracy of polyp size assessment in colonoscopy.

Main Methods:

• A novel method was developed to estimate a polyp's three-dimensional spatial information, including localization and depth.
• The combination of location and depth data was used to represent the polyp's 3D shape.
• The method was evaluated using 787 polyps of both protruded and flat types.

Main Results:

• The proposed method achieved superior classification accuracy compared to state-of-the-art image classification techniques.
• Sequential voting post-processing enhanced classification accuracy, reaching 0.81 for 1-9 mm polyps and 0.88 for larger polyps.
• Qualitative analysis confirmed the critical role of polyp localization in size classification.

Conclusions:

• A binary polyp-size classification method utilizing estimated 3D polyp shape was successfully developed.
• The method demonstrated accurate classification for both protruded and flat polyps.
• Accurate classification was achieved even for flat polyps with ambiguous boundaries.