Automated classification of nasal polyps in endoscopy video-frames using handcrafted and CNN features

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Otolaryngology

Background:

• Nasal polyps are common growths that can lead to allergic rhinitis, sinus infections, and asthma.
• Accurate and timely diagnosis of nasal polyps is crucial for effective patient management.
• Existing diagnostic methods may benefit from advanced computational tools for improved accuracy and efficiency.

Purpose of the Study:

• To design a reliable rhinology assistance system for recognizing nasal polyps in endoscopic videos.
• To introduce and benchmark a novel dataset (NP-80) for nasal polyp classification.
• To develop and evaluate deep learning models for automated nasal polyp detection.

Main Methods:

• A new dataset, NP-80, comprising high-quality endoscopy video frames from 80 participants, was created.
• Machine learning and deep learning classifiers, including KNN, SVM, RF, DT, and CNN, were benchmarked.
• Handcrafted features (LBP, HOG) and CNN features were compared for polyp recognition performance.

Main Results:

• The proposed Convolutional Neural Network (CNN) classifier achieved the highest accuracy of 98.3%.
• Precision and recall rates reached 99% and 98%, respectively, demonstrating high performance.
• Classifiers using handcrafted features showed lower performance, with a best accuracy of 96.3%.

Conclusions:

• Deep learning models, particularly CNNs, show significant promise for accurate nasal polyp classification from endoscopic images.
• The developed system can serve as a valuable decision support tool for otolaryngologists.
• The publicly released NP-80 dataset will facilitate further research in computational rhinology.