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

Area of Science:

• Medical Physics
• Artificial Intelligence
• Radiotherapy

Background:

• Radiation therapy planning is complex, requiring dose optimization for targets and sparing of normal tissues.
• Knowledge-based planning (KBP) is increasingly demanded to improve efficiency and quality in treatment planning.
• Ensemble learning offers potential for KBP but faces challenges with learner diversity, accuracy, and model complexity.

Purpose of the Study:

• To introduce LENAS, a novel learning-based ensemble approach for 3D radiotherapy dose prediction.
• To integrate neural architecture search and knowledge distillation to create efficient and effective planning models.
• To address the complexity and computational overhead associated with traditional model ensembles.

Main Methods:

• Utilized exhaustive neural architecture search to identify diverse and high-performing base models.
• Employed a teacher-student paradigm for knowledge distillation, using diverse ensemble outputs as supervisory signals.
• Designed a hybrid loss function to preserve high-level semantic information during student network training.

Main Results:

• LENAS was evaluated on the OpenKBP and AIMIS datasets.
• Experimental results demonstrated the effectiveness and superiority of the proposed method compared to state-of-the-art approaches.
• The approach successfully mitigated the complexity of model ensembles while maintaining performance.

Conclusions:

• LENAS offers a promising solution for enhancing 3D radiotherapy dose prediction through efficient knowledge-based planning.
• The integration of neural architecture search and knowledge distillation effectively balances performance, diversity, and model complexity.
• This work advances the application of AI in radiotherapy, paving the way for more streamlined and accurate treatment planning.