Towards Boundary More Precise Detection: Surrounding-to-aggregating Deep Learning in Videoscope Imaging.

The assessment of early laryngeal cancer and pre-neoplastic lesions is subjective and depends on doctors' experience, leading to missed diagnoses in primary institutions. Our objective was to develop and validate a deep learning algorithm for the real-time identification of early laryngeal cancer and pre-neoplastic lesions, aiming to enhance diagnostic accuracy. The challenge observed in the domain of deep learning arises from overlooking contextual information. In response, we introduce in this paper a learning methodology that advances from acknowledging the surrounding context to integrating it, providing a resolution to this problem. Initially, we introduce side-aware features to capture relevant characteristics. Subsequently, we employ a rectangular selection technique for accurately determining regions of interest. To assess the effectiveness of our approach in object detection, we perform evaluations on a clinical dataset. Our deep learning approach exhibits robust performance in discriminating cancer. The images were randomly divided into training (80%), testing (10%), and validation (10%) sets. The testing was performed on a laryngoscope dataset consisting of 1123 samples. When compared with other advanced detection models, our methodology surpassed them, demonstrating superior results in laryngoscope detection, including mAP, accuracy, recall, and F1 score. In this study, we identified a learning method conducive to polyp detection in video laryngoscopy under both white-light and narrow-band imaging. The promising detection performance holds the potential to improve diagnostic proficiency and decrease the likelihood of missed diagnoses among primary otolaryngologists. [ABSTRACT FROM AUTHOR]