Medical Image Classification Based on Information Interaction Perception Mechanism.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Gastroenterology
• Computational Pathology

Background:

• Colorectal cancer (CRC) develops from adenomatous polyps, which can become malignant and spread, leading to fatal complications.
• Diagnostic accuracy in colonoscopy is challenged by factors like operator experience and visual fatigue.
• There is a need for automated systems to support medical imaging personnel in polyp detection.

Purpose of the Study:

• To propose an automated network model for colonic polyp detection using colonoscopy images.
• To enhance the accuracy of polyp classification and reduce computational costs.
• To address the challenge of detecting polyps with unnoticeable surface textures.

Main Methods:

• Development of a Channel Information Interaction Perception (CIIP) module to capture subtle polyp textures.
• Introduction of the Information Interaction Perception Network (IIP-Net) incorporating the CIIP module.
• Evaluation of IIP-Net using three classification structures: fully connected (FC), global average pooling fully connected (GAP-FC), and convolution global average pooling (C-GAP).

Main Results:

• The IIP-NET54-GAP-FC module demonstrated a high overall accuracy of 99.59% in detecting colonic polyps.
• The specific accuracy for colonic polyp detection reached 99.40%.
• The proposed IIP-NET54-GAP-FC model exhibited superior performance compared to other evaluated configurations.

Conclusions:

• The IIP-Net, particularly the IIP-NET54-GAP-FC configuration, is a highly accurate and effective tool for colonic polyp detection in colonoscopy images.
• This AI model shows promise in assisting medical professionals, potentially improving diagnostic efficiency and patient outcomes in colorectal cancer screening.
• The CIIP module effectively addresses the challenge of detecting polyps with subtle surface textures.