Spatio-Temporal Feature Transformation Based Polyp Recognition for Automatic Detection: Higher Accuracy than Novice...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Artificial intelligence (AI) shows promise for enhancing colonoscopy quality.
• Developing AI tools for polyp detection is crucial for improving patient outcomes.

Purpose of the Study:

• To develop a Short-Time Fourier Transform (STFT) based colon polyp detection model.
• To evaluate the model's performance against endoscopists in a randomized experiment.

Main Methods:

• Colonoscopy videos from January 2018 to November 2022 were utilized.
• A dataset of 1500 colonoscopy images and 1200 polyp images was randomly selected for testing.
• The STFT model's recognition results were compared with endoscopists of varying experience levels.

Main Results:

• The STFT model achieved higher accuracy and specificity than less experienced endoscopists.
• Model sensitivity (0.904) surpassed all endoscopist groups (P < 0.05).
• The model demonstrated superior accuracy for small polyps (≤ 0.5 cm and 0.6-1.0 cm) compared to novice endoscopists.

Conclusions:

• The STFT-based AI model accurately detects colon polyps in videos.
• The model is particularly effective in identifying small polyps (≤ 0.5 cm).
• This AI tool has significant potential to improve colonoscopy diagnostic accuracy.