A Novel Framework of Manifold Learning Cascade-Clustering for the Informative Frame Selection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Narrow band imaging (NBI) is crucial for early laryngeal cancer detection.
• Many NBI images are uninformative (blurred, reflections, underexposed), hindering accurate computer-aided diagnosis.
• Manual review of NBI frames is time-consuming and subjective.

Purpose of the Study:

• To develop an unsupervised scheme for automatically identifying informative NBI frames.
• To improve the efficiency and accuracy of computer-aided diagnosis for laryngeal cancer.
• To eliminate the need for tedious manual labeling of NBI frames.

Main Methods:

• Feature embedding extraction using a VGG16 neural network.
• Dimensionality reduction using Uniform Manifold Approximation and Projection (UMAP).
• Automatic cluster labeling and Bayesian optimization for cost function refinement.

Main Results:

• The unsupervised scheme achieved state-of-the-art performance, outperforming baselines by 12%.
• The proposed method demonstrated a high overall median recall of 96%.
• UMAP effectively distinguished feature embeddings in a lower-dimensional space.

Conclusions:

• The novel unsupervised scheme is effective and robust for detecting informative NBI frames.
• This approach enhances computer-aided diagnosis accuracy and speed for laryngeal cancer surveillance.
• The findings suggest data patterns can enable flexible algorithms without manual labeling.