Attention-guided deep framework for polyp localization and subsequent classification via polyp local and Siamese...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Colorectal cancer (CRC) is a major global health concern.
• Accurate polyp detection and classification are crucial for early diagnosis and treatment.
• Existing methods may lack the precision for real-time colonoscopy analysis.

Purpose of the Study:

• To develop an automated deep learning model for polyp detection, localization, and classification in colonoscopy videos.
• To enhance polyp localization using spatial and channel attention mechanisms within the YOLOv4 architecture.
• To classify polyps as adenoma or non-adenoma by fusing deep and handcrafted features.

Main Methods:

• Utilized the YOLOv4 deep learning model with spatial and channel attention blocks for polyp localization.
• Employed Pyramid Histogram of Oriented Gradients (PHOG) for local shape feature extraction.
• Integrated a triplet Siamese network for learning intra-polyp discriminating features.
• Fused deep and handcrafted features for polyp classification.

Main Results:

• Achieved high polyp localization performance with average precision of 0.8971 (Kvasir-SEG) and 0.9171 (SUN).
• Demonstrated robust classification accuracy of 96.66% for polyp types (adenoma/non-adenoma).
• Attained an F1 score of 96.54% for the classification framework, validating its effectiveness.

Conclusions:

• The proposed automated technique effectively detects, localizes, and classifies polyps in colonoscopy videos.
• The fusion of YOLOv4, attention mechanisms, PHOG, and Siamese networks offers a robust diagnostic tool.
• This method shows significant potential for real-time and offline analysis in colonoscopy, aiding in colorectal cancer prevention.