LENAS: Learning-based Neural Architecture Search and Ensemble for 3D Radiotherapy Dose Prediction

Radiation therapy treatment planning is a complex process, as the target dose prescription and normal tissue sparing are conflicting objectives. In order to reduce human planning time efforts and improve the quality of treatment planning, knowledge-based planning (KBP) is in high demand. In this study, we propose a novel learning-based ensemble approach, named LENAS, which integrates neural architecture search (NAS) with knowledge distillation for 3D radiotherapy dose prediction. Specifically, the prediction network first exhaustively searches each block from an enormous architecture space. Then, multiple architectures with promising performance and a large diversity are selected. To reduce the inference time, we adopt the teacher-student paradigm by treating the combination of diverse outputs from multiple learned networks as supervisions to guide the student network training. In addition, we apply adversarial learning to optimize the student network to recover the knowledge in teacher networks. To the best of our knowledge, this is the first attempt to investigate NAS and knowledge distillation in ensemble learning, especially in the field of medical image analysis. The proposed method has been evaluated on two public datasets, i.e., the OpenKBP and AIMIS dataset. Extensive experimental results demonstrate the effectiveness of our method and its superior performance to the state-of-the-art methods. In addition, several in-depth analysis and empirical guidelines are derived for ensemble learning.