Medical Imaging Fusing Vision Transformer: Laryngeal Cancer Screening with Explanation

Laryngeal cancer diagnosis via Narrow Band Imaging (NBI) endoscopy is highly dependent on clinician expertise, making it prone to interobserver variability and time-consuming workflows. This study investigates whether a deep learning framework can automate the classification of laryngeal lesions while providing interpretable visual feedback to assist clinicians in their decision-making process.

The researchers implemented a dual-component framework. First, they utilized a Swin Transformer—a hierarchical vision transformer architecture—to classify NBI endoscopic images as either benign or malignant. Second, they integrated MedSAM, a state-of-the-art medical image segmentation foundation model, to generate visual masks that highlight the specific pathological regions influencing the classification. The model was trained and tested on the public CE-NBI dataset, which contains over 11,000 images from 210 patients.

The proposed Swin Transformer model achieved an overall classification accuracy of 82.33% and a weighted F1-score of 82.72%. The integration of MedSAM successfully provided visual explanations by segmenting suspicious areas, which the authors argue bridges the gap between black-box AI predictions and clinical interpretability. The model demonstrated strong precision for malignant cases, suggesting potential utility in reducing unnecessary biopsies.

By fusing classification with segmentation-based explanation, this work addresses a critical barrier to the clinical adoption of AI in oncology: the lack of transparency. Providing clinicians with both a diagnostic prediction and a visual indication of the relevant pathological features may improve trust and facilitate the integration of AI-assisted screening into routine laryngoscopic examinations.