深度学习预测前列腺癌自动治疗计划中每个控制点的监测单位时单模型与多模型的比较

Single- versus multi-model in the deep learning prediction of monitor units per control point for automated treatment planning in prostate cancer.

作者信息

Gaudreault Mathieu, McIntosh Lachlan, Woodford Katrina, Li Jason, Harden Susan, Porceddu Sandro, Panettieri Vanessa, Hardcastle Nicholas

机构信息

Peter MacCallum Cancer Centre, Melbourne, Victoria, Australia.

Sir Peter MacCallum Department of Oncology, The University of Melbourne, Melbourne, Victoria, Australia.

出版信息

J Appl Clin Med Phys. 2025 Sep;26(9):e70229. doi: 10.1002/acm2.70229.

DOI:10.1002/acm2.70229

PMID:40883098

原文链接:https://pmc.ncbi.nlm.nih.gov/articles/PMC12396887/

Abstract

BACKGROUND

In contemporary radiation therapy, the radiation is modulated to conform the prescription dose to the tumor and spare organs at risk. The modulation results from a complex mathematical calculation that requires several iterations to reach a satisfactory solution, delaying treatment. The monitor units (MU) per control point (CP) control the dose magnitude and may be predicted by deep learning, a type of artificial intelligence (AI).

PURPOSE

To introduce deep learning methods to predict the MU per CP in the context of AI volumetric modulated arc therapy (VMAT) treatment plan prediction for prostate cancer.

METHODS

Patients treated for prostate cancer with 60 Gy in 20 fractions between 01/2019 and 06/2024 were considered for inclusion. Two approaches were considered: a single-model approach, trained on all samples, and a multi-model approach, with separate models trained by CP. The inputs were either the three-dimensional (3D) dose per CP (3D single-model / 3D multi-model) or the two-dimensional (2D) average dose intensity projection per CP (2D single-model / 2D multi-model). The outputs were the MU per CP, which were converted to meterset weight per CP and MU per beam to create an AI-Radiation Therapy Plan (AI-RTPlan) with other clinical parameters retained. Clinical goals achieved with the calculated dose distribution from the AI-RTPlan and clinical plan were compared.

RESULTS

The cohort was split into 220/40/42 homogeneous plans in the training/validation/testing dataset. Relative to the clinical case, the errors in meterset weight per CP were mean ± SD = -0.4 ± 3.8%/-0.2 ± 4.8% in 2D/3D single-model and 0.01 ± 3.9%/-0.1 ± 5.0% in 2D/3D multi-model. The errors in MU per beam were -0.9 ± 5.5%/-1.2 ± 4.5% in 2D/3D single-model and 0.4 ± 4.8%/0.5 ± 5.2% in 2D/3D multi-model. In 2D/3D models, at least 93%/81% of patients had the same or more clinical goals achieved with AI-RTPlans.

CONCLUSIONS

Accurate prediction of MU per CP is feasible in VMAT prostate cancer treatment.

摘要

背景

在当代放射治疗中，辐射被调制以使处方剂量与肿瘤形状相符，并保护危及器官。这种调制源于复杂的数学计算，需要多次迭代才能得到满意的解决方案，从而延迟了治疗。每个控制点（CP）的监测单位（MU）控制着剂量大小，并且可以通过深度学习（一种人工智能（AI）类型）进行预测。

目的

在人工智能容积调强弧形放疗（VMAT）治疗前列腺癌的计划预测背景下，引入深度学习方法来预测每个CP的MU。

方法

纳入2019年1月至2024年6月期间接受20次分割、总剂量60 Gy的前列腺癌治疗患者。考虑了两种方法：一种是单模型方法，在所有样本上进行训练；另一种是多模型方法，按CP分别训练模型。输入数据要么是每个CP的三维（3D）剂量（3D单模型/3D多模型），要么是每个CP的二维（2D）平均剂量强度投影（2D单模型/2D多模型）。输出是每个CP的MU，将其转换为每个CP的跳数权重和每个射野的MU，以创建一个保留其他临床参数的人工智能放射治疗计划（AI-RTPlan）。比较了通过AI-RTPlan和临床计划计算的剂量分布所实现的临床目标。

结果

在训练/验证/测试数据集中，队列被分为220/40/42个同质计划。相对于临床病例，2D/3D单模型中每个CP的跳数权重误差为均值±标准差=-0.4±3.8%/-0.2±4.8%，2D/3D多模型中为0.01±3.9%/-0.1±5.0%。每个射野的MU误差在2D/3D单模型中为-0.9±5.5%/-1.2±4.5%，在2D/3D多模型中为0.4±4.8%/0.5±5.2%。在2D/3D模型中，至少93%/81%的患者通过AI-RTPlan实现了相同或更多的临床目标。