Single-View Contrastive Learning for Laryngeal Leukoplakia Classification With NBI Laryngoscopy Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Laryngeal cancer, a prevalent upper respiratory malignancy, necessitates early and accurate diagnosis for improved patient outcomes.
• Laryngoscopy with Narrow Band Imaging (NBI) aids endoscopists, but fine classification of laryngeal leukoplakia from NBI images remains a challenge for computer-aided diagnosis.
• Developing robust AI models is crucial for enhancing diagnostic accuracy in challenging cases.

Purpose of the Study:

• To develop a computer-aided diagnosis system for fine classification of laryngeal leukoplakia using NBI images.
• To address the challenge of small sample sizes in medical image datasets through effective learning strategies.
• To improve the accuracy and efficiency of laryngeal cancer diagnosis.

Main Methods:

• A single-view contrastive learning network was proposed, incorporating lesion localization and pseudo-labeling for unlabeled data.
• A backbone network was pretrained on original NBI images, followed by patch generation using an attention-guided network to augment samples.
• Joint training of the backbone and contrastive learning networks was performed using a combined contrastive and cross-entropy loss function.

Main Results:

• The proposed model achieved a high accuracy of 96.12% on the NBI dataset, surpassing current mainstream models.
• The model demonstrated high specificity and sensitivity in classifying six categories: normal tissue, inflammatory keratosis, mild, moderate, severe dysplasia, and squamous cell carcinoma.
• Experimental validation confirmed the model's effectiveness in fine classification tasks.

Conclusions:

• The developed single-view contrastive learning network offers a promising approach for accurate and efficient fine classification of laryngeal leukoplakia.
• The model's superior performance highlights its potential to aid clinicians in the early diagnosis of laryngeal cancer.
• The availability of the code facilitates further research and development in AI-driven medical diagnostics.