Artificial Intelligence in the Diagnosis of Laryngeal Cancer Based on Endoscopic Images: A Comprehensive Narrative Review

Background: Laryngeal cancer represents a significant global health burden with early detection being crucial for improved patient outcomes. The integration of artificial intelligence (AI) into medical diagnostics has shown tremendous potential in enhancing the accuracy and efficiency of laryngeal cancer detection through endoscopic image analysis. Objective: This narrative review comprehensively examines the current state of AI applications in laryngeal cancer diagnosis using endoscopic imaging, focusing on deep learning methodologies, diagnostic performance, clinical validation and future perspectives. Methods: A systematic analysis of peer-reviewed literature was conducted, examining studies published between 2016 and 2025 that investigated AI-based approaches for laryngeal cancer detection and classification using endoscopic images. The review encompassed various AI techniques including convolutional neural networks (CNNs), transfer learning, multimodal approaches and novel architectures applied to white light imaging, narrow-band imaging and other advanced endoscopic modalities. Results: Current AI systems demonstrate remarkable diagnostic accuracy with sensitivity ranging from 71-98% and specificity from 86-98% across different studies and methodologies. Deep learning approaches, particularly CNNs such as ResNet, DenseNet, and YOLO architectures, have shown performance comparable to or exceeding that of experienced clinicians. Multimodal AI systems combining white light and narrow-band imaging data achieve superior performance compared to single-modality approaches. Real-time processing capabilities with inference times as low as 0.01-0.03 seconds per image enable practical clinical implementation. Conclusions: AI-assisted endoscopic diagnosis represents a transformative technology for laryngeal cancer detection, offering the potential to standardize diagnostic protocols, reduce inter-observer variability and improve access to expert-level diagnostic capabilities globally.