THTL: An Effective Two-Step Heterogeneous Transfer Learning Framework for Early Laryngeal Cancer Identification

Early and accurate identification of benign or malignant laryngeal lesions is crucial for early intervention in laryngeal cancer. To achieve this goal, we propose a two-step heterogeneous transfer learning (THTL) framework based on traditional transfer learning. This framework utilizes heterogeneous medical images from coronary angiography as an intermediate domain, which while differing in semantics and imaging instruments, shares vascular features with the target domain (laryngeal blood vessel images) to a certain extent. We verified our proposed THTL on three different scales of target domains: two extremely challenging ones using only 2% and 10% of the original laryngeal blood vessel dataset for training and validation, and the other larger one using 90%. Compared to traditional transfer learning, THTL achieves better results across all scales. Specifically, THTL improves accuracy, malignant precision, and malignant recall by 3.45%, 13.1%, and 18.73%, and by 9.93%, 11.62%, and 36.76%, respectively, for two limited data scenarios, and exceeds the state-of-the-art accuracy in the target domain scale of 90%. Additionally, to gain deeper insights, a number of comprehensive experiments are conducted on fine-tuning techniques, domain color impacts, and the intermediate domain’s size and complexity, providing valuable information for subsequent studies in applying transfer learning for early laryngeal cancer identification.