Transfer learning for informative-frame selection in laryngoscopic videos through learned features.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Oncology diagnostics

Background:

• Narrow-band imaging (NBI) laryngoscopy aids in laryngeal cancer diagnosis but generates large datasets.
• Manual review of NBI laryngoscopy videos is time-consuming and may miss critical frames.
• Reducing data volume is crucial for efficient and accurate laryngeal cancer screening.

Purpose of the Study:

• To develop a deep learning strategy for automatic selection of informative frames from NBI laryngoscopy videos.
• To reduce the computational load and improve diagnostic efficiency in laryngeal cancer detection.
• To leverage transfer learning for feature extraction in medical video analysis.

Main Methods:

• Utilized transfer learning with six pre-trained convolutional neural networks (CNNs) for feature extraction from NBI frames.
• Extracted features from the NBI-InfFrames dataset.
• Employed Support Vector Machines (SVMs) and CNN-based classifiers to distinguish informative frames from uninformative ones (blurred, saliva, underexposed).

Main Results:

• The VGG 16 model achieved the highest performance in feature extraction.
• SVM classification with VGG 16 features yielded a recall of 0.97 for informative frames.
• CNN-based classification with VGG 16 features resulted in a recall of 0.98 for informative frames.

Conclusions:

• The proposed deep learning strategy effectively automates informative frame selection in laryngoscopic videos.
• Transfer learning demonstrates significant potential for enhancing medical image analysis tasks.
• This approach offers a valuable tool for improving the efficiency and accuracy of laryngeal cancer diagnosis.