Dynamic Ensemble Transfer Learning with Multi-view Ultrasonography for Improving Thyroid Cancer Diagnostic Reliability.

Diagnostic decision-making requires the integration of relevant facts and clinician experience. Incorporating the clinical experience from diverse backgrounds is beneficial in a multi-disciplinary model to mitigate uncertainties aroused by incomplete mastery of knowledge. However, current computer-aided diagnostic systems are generally designed using unitary datasets and are challenging to adapt to diverse institutions, leading to the limited reliability of the generated decisions. Accordingly, this study proposes a dynamic ensemble transfer learning-based system that simulates such diversity in its training and structure by integrating knowledge and data. The approach consists of a self-directed model selection scheme, a dynamic weighting mechanism, and a unified weighted ensemble averaging model, tailored for reliable diagnostic decision-making. This study adopts the most rapidly rising malignancy worldwide, thyroid cancer, for evaluation. Two multi-view thyroid ultrasonography datasets with matching tissue diagnosis from over 700 cross-national patients are used to pre-train the individual networks. The learnt knowledge is then transferred to the weighted ensemble averaging model through the dynamic weighting mechanism. The fine-tuned ensemble model is evaluated using an external set of thyroid nodules with radiological risk of malignancy based on the Thyroid Imaging Reporting and Data System. Further, we alter the datasets through up-sampling and down-sampling to evaluate the ensemble model's generalization. Extensive experiments demonstrate that the proposed ensemble model yields promising performance with an area under the curve value between 0.87 and 0.93 under diversified strategies. Benchmarking results show the proposed approach surpasses existing studies and improves diagnostic reliability in thyroid cancer care while guiding subsequent management options.