Feature extraction and pattern classification of colorectal polyps in colonoscopic imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Accurate classification of colorectal polyps is crucial for effective cancer prevention.
• Traditional visual inspection by physicians can be subjective and prone to error.
• Computer-aided diagnostic (CAD) systems offer potential for improved accuracy and consistency.

Purpose of the Study:

• To develop and evaluate a CAD system for classifying colorectal polyps from colonoscopic images.
• To compare the diagnostic performance of the CAD system against expert visual inspection.

Main Methods:

• Image enhancement using Principal Component Transform (PCT).
• Feature extraction including texture, spatial, and spectral domains.
• Feature selection using Sequential Forward Selection (SFS) and Sequential Floating Forward Selection (SFFS).
• Classification using Support Vector Machines (SVMs).

Main Results:

• The CAD system achieved an Az value of 88.7% with all features.
• SFFS-selected features improved classification accuracy to 93.1% (Az value) and reduced feature dimensions by 73.8%.
• The final system achieved 96% accuracy in polyp type classification, surpassing the 85% accuracy of experienced physicians.

Conclusions:

• The proposed CAD system significantly improves the accuracy of colorectal polyp diagnosis.
• The SFFS method is effective for feature selection, enhancing both classification performance and efficiency.
• This AI-powered system has the potential to enhance the quality and reliability of colonoscopic diagnoses.