Transfer Learning on a Deep Learning Approach to Optimize Laryngeal Cancer Detection on Endoscopic Images

Laryngeal cancer is a significant oncologic condition, with its prognosis heavily influenced by the implementation of timely preventive and diagnostic measures. The outlook for laryngeal cancer is generally better when diagnosed early. As the booming of AI, a deep learning approach to assistant diagnostic may provide a solution. But to apply the deep learning for laryngeal cancer, the rare data is always its suffering. To solve this bottleneck, we investigated different transfer learning schemes on a deep learning approach to optimize laryngeal cancer detection. The experiment is set on 11,144 images extracted from CE-NBI video recordings. The results are analyzed by metrics (precision, recall, and F1) in this public hospital dataset. The F1 could reach 0.84. It indicates that the approach has the ability to generalize in distinguishing benign and malignant with reliable balanced accuracy. It has good potential as a preliminary diagnosis assistant in the detection of laryngeal cancer.