Learning-based classification of informative laryngoscopic frames.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Early diagnosis of laryngeal cancer is crucial for reducing patient morbidity.
• Narrow-band imaging (NBI) endoscopy aids in screening but generates large datasets.
• Current methods require extensive review of endoscopic video data.

Purpose of the Study:

• To develop an automated strategy for selecting informative endoscopic video frames.
• To reduce the volume of data requiring manual review for diagnosis.
• To enhance the performance of computer-assisted diagnosis systems.

Main Methods:

• A novel classification method using intensity, keypoint, and spatial content features.
• Support vector machines with radial basis function and one-versus-one scheme for frame classification.
• Classification categories include informative, blurred, saliva, specular reflections, and underexposed frames.

Main Results:

• Achieved 91% recall for identifying informative frames on a balanced dataset of 720 images.
• Significantly outperformed three state-of-the-art methods in frame classification.
• Demonstrated high performance in distinguishing informative frames from artifacts.

Conclusions:

• The proposed approach is a valuable tool for automated informative frame selection.
• Potential applications include computer-assisted diagnosis and endoscopic view enhancement.
• This method can streamline the analysis of NBI endoscopic videos.