Enhanced WGAN Model for Diagnosing Laryngeal Carcinoma

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Laryngeal cancer diagnosis relies on endoscopic imaging, often requiring manual segmentation.
• Accurate pixel-based segmentation of lesions is crucial for early detection and treatment planning.
• Existing segmentation models face challenges with complex lesion characteristics and data variability.

Purpose of the Study:

• To develop an advanced U-Net model for automated pixel-based segmentation and classification of laryngeal lesions.
• To improve the accuracy and robustness of lesion detection in endoscopic images.
• To enhance early cancer diagnosis and reduce diagnostic errors.

Main Methods:

• Modification of the U-Net architecture with five-level encoders/decoders and an autoencoder layer.
• Integration of a Wasserstein Generative Adversarial Network (WGAN) to stabilize training and prevent mode collapse.
• Utilized a dataset of 8171 laryngeal endoscopic images with multi-class polygon annotations.

Main Results:

• The enhanced U-Net model achieved an overall accuracy of 99% in lesion classification.
• Cancers were detected with notably high accuracy, while benign tumors showed lower detection rates.
• Evaluation metrics, including F1 score and intersection over union, confirmed model efficacy.

Conclusions:

• The enhanced U-Net model demonstrates significant potential for accurate laryngeal lesion segmentation and classification.
• This AI-driven approach can improve early cancer detection rates and reduce diagnostic inaccuracies.
• The model offers a promising tool for enhancing clinical decision-making in laryngology.