A hybrid meta-heuristic framework with ensemble deep learning for multi-functional simultaneous optimized automatic intensity-modulated radiotherapy planning

• Published in: Expert systems with applications 2025-01, Vol.260, p.125485, Article 125485
• Main Authors: Yang, Xiaoyu, Li, Shuzhou, Shao, Qigang, Tang, Du, Peng, Zhao, Cao, Ying, Yang, Zhen, Zhao, Yuqian
• Format: Article
• Language: English
• Subjects: Automatic treatment planning for radiotherapy; Deep learning; Delivery accuracy enhancement; Multi-functional simultaneous optimization
• ISSN: 0957-4174
• DOI: 10.1016/j.eswa.2024.125485

Intensity-modulated radiotherapy (IMRT) is one of the main treatments for patients with cancer, and its treatment planning holds significant importance. Compared to manual treatment planning, automatic treatment planning (ATP) is expected to improve plan quality and efficiency, which is of great clinical importance. However, treatment planning is an intractable “black art” that calls for intuition and heuristics from medical dosimetrists. There are two main issues with the current ATP methods: (i) they rely strongly on prior clinical experience and are not sufficiently scalable, and (ii) they use a single-functional optimization concept that ignores delivery accuracy information, which lowers the delivery accuracy of the generated ATP plans. To address these issues, this paper presents a novel multi-functional simultaneous optimized automatic treatment planning (MFSO-ATP) method. First, to address the issues of high reliance and insufficient scalability, an evolutionary strategy (ES) and simulated annealing (SA)-based hybrid meta-heuristic trial-and-error (ESSA-HMTE) framework is designed to automatically adjust optimization parameters without human intervention. Second, to address the issue of delivery accuracy deterioration, an ensemble deep learning-based gamma passing rate (GPR) prediction (EnDL-GPR) framework was constructed and trained to predict the plan delivery accuracy a priori. The ESSA-HMTE framework effectively fuses dosimetric quality information and delivery accuracy a priori information to simultaneously optimize the two, ensuring delivery accuracy while achieving automatic planning. We tested the MFSO-ATP method using multiple diseases with high inter-tumor heterogeneity. The results showed that the proposed method can improve the quality and delivery accuracy of IMRT plans, reduce the percentage of quality assurance failures, is highly scalable, and has good clinical application prospects.