Robust endocytoscopic image classification based on higher-order symmetric tensor analysis and multi-scale...

Area of Science:

• Medical imaging
• Computer-aided diagnosis
• Endoscopy

Background:

• Endocytoscopy offers cell-level observation during colonoscopy, potentially improving cost-effectiveness.
• Accurate endocytoscopic diagnosis requires significant physician expertise, posing a challenge for widespread adoption.
• Computer-aided diagnosis (CAD) systems can enhance accuracy and reduce interobserver variability in medical image interpretation.

Purpose of the Study:

• To develop a robust endocytoscopic (EC) image classification method for a computer-aided diagnosis (CAD) system.
• To address the challenges of expert knowledge and interobserver variability in EC image diagnosis.
• To improve the practical utility and accuracy of endocytoscopy in colonoscopy.

Main Methods:

• Proposed a novel feature extraction method using higher-order symmetric tensor analysis for multi-scale topological statistics.
• Integrated the novel feature extraction with endocytoscopic (EC) image classification.
• Evaluated the method's classification accuracy against three deep learning approaches using a large, multi-hospital dataset of approximately 55,000 images from over 3800 patients.

Main Results:

• The proposed method achieved an average classification accuracy of 90% across four hospitals.
• The method demonstrated robustness against variations in pit patterns, including color, contrast, shape, and hospital-specific differences.
• With a rejection option, the proposed method achieved expert-level classification accuracy.

Conclusions:

• A robust EC image classification method with novel feature extraction was developed.
• The method exhibits strong generalization ability, making it suitable for constructing practical CAD systems.
• This advancement supports the integration of endocytoscopy into routine colonoscopy, enhancing diagnostic capabilities.