AI-Enabled Classification of Head and Neck Tumors using Convolutional Neural Networks with Dynamic Optical Contrast Imaging

BackgroundRecent advances in artificial intelligence (AI) in the field of imaging have resulted in new opportunities for automated tumor detection and margin assessment. In particular, AI deep learning techniques such as the Convolutional Neural Network (CNN) have greatly advanced the field of computer vision. Here we introduce the application of a CNN model for use with Dynamic Optical Contrast Imaging (DOCI), an imaging technique developed by our group that creates a unique molecular signature on tissue targets by obtaining the autofluorescence decay of several spectral bands in the UV-Vis range. Methods21 patients undergoing surgical resection for tonsillar squamous cell carcinoma (SCC) were identified on a prospective basis. DOCI images were analyzed and compared to the pathology results as ground truth. A CNN model was used to segment sections of DOCI images and provide a percentage chance of tumor presence, allowing for automated tumor margin delineation without a-priori knowledge of the tumor tissue composition. ResultsCNN outputs yielded a 99.98% confidence in classifying non-tumor tissue and 76.02% confidence in classifying tumor tissue. ConclusionsOur results indicate that a CNN-based classification model for DOCI allows for real-time analysis of tissue, providing improved sensitivity and accuracy of determining true margins and thus enabling the head and neck cancer surgeon to save healthy tissue and improve patient outcomes.