Reliable lymph node metastasis prediction in head neck cancer through automated multi-objective model

Lymph node metastasis (LNM) plays an important role for accurately diagnosing and treating the patients with head neck cancer. Positron emission tomography (PET) and computed tomography (CT) are two primary imaging modalities used for identifying LNM status. However, the uncertainty of LNM may exist especially for reactive or small nodes. Furthermore, identifying the LNM on PET or CT is greatly dependent on the physician's experience. Therefore, developing a reliable and automatic model is essential for accurately identifying LNM. Multi-objective models have shown promising predictive results by considering different objectives such as sensitivity and specificity. However, most multi-objective models need to choose an optimal model manually. In this work, we proposed an automated multi-objective learning model (AutoMO) for predicting LNM reliably. Instead of picking one optimal model, all the Pareto-optimal models with the calculated relative weights are used in AutoMO. Then the evidential reasoning (ER) approach is used for fusing the output probability for obtaining more reliable results than traditional fusion method. We built three models for PET, CT and PETCT and the results showed that PETCT outperformed two single modality based models. The comparative study demonstrated that AutoMO obtained better performance than current available multi-objective and deep learning methods, and more reliable results can be acquired when using ER fusion.