DEVELOPMENT OF ARTIFICIAL INTELLIGENCE- ENHANCED OPTICAL BIOPSY TOOLS FOR ENDOSCOPIC EVALUATION OF LARYNGEAL TUMORS

Head and neck squamous cell carcinoma, particularly laryngeal cancer, remains a significant global health challenge. Early detection is crucial for improving survival and preserving organ function, yet endoscopic evaluation of laryngeal lesions remains highly operator-dependent and subject to inter-observer variability. Advances in artificial intelligence (AI), especially deep learning applied to medical imaging, offer the potential to enhance the accuracy, objectivity, and efficiency of endoscopic assessment. This thesis investigates the development and clinical feasibility of AI-enhanced optical biopsy tools for the automated analysis of laryngeal endoscopy. Multiple deep learning models were designed and evaluated across different tasks relevant to clinical practice. First, an algorithm for automatic informative frame selection was developed to identify diagnostically relevant images within laryngoscopic videos, enabling real-time feedback and optimized data extraction. Second, computer-aided detection models were implemented to automatically identify suspicious laryngeal lesions in endoscopic images. Third, diagnostic classification models were trained to differentiate between benign and malignant lesions, with performance compared against clinicians and large language models. Further investigations focused on the automatic segmentation of tumor boundaries using convolutional neural networks, enabling precise delineation of lesion extent in both white light and narrow-band imaging. Finally, a deep learning framework was developed for quantitative assessment of laryngeal motility, providing an objective tool for evaluating vocal fold function in clinical settings. Across these studies, models were trained and validated on multicenter datasets and tested on external cohorts to assess generalizability. Results demonstrate that AI systems can achieve high accuracy and real-time performance in multiple tasks of laryngeal endoscopic analysis, with performances comparable to or exceeding human raters in specific scenarios. Overall, this work highlights the potential of AI-assisted endoscopy to support clinicians, reduce diagnostic variability, and move toward real-time computer-assisted optical biopsy in laryngeal cancer evaluation.