Predicting anatomical variations in radiotherapy with a vector quantized variational autoencoder generative model

IF 3.2 2区医学 Q1 RADIOLOGY, NUCLEAR MEDICINE & MEDICAL IMAGING

Medical physics Pub Date : 2025-09-12 DOI:10.1002/mp.18120

Yue Zou, Zhenhao Li, Menghan Zhang, Ziwei Li, Xiaojie Yin, Long Yang, Weigang Hu, Jiazhou Wang

{"title":"Predicting anatomical variations in radiotherapy with a vector quantized variational autoencoder generative model","authors":"Yue Zou, Zhenhao Li, Menghan Zhang, Ziwei Li, Xiaojie Yin, Long Yang, Weigang Hu, Jiazhou Wang","doi":"10.1002/mp.18120","DOIUrl":null,"url":null,"abstract":"<div>\n \n \n <section>\n \n <h3> Background</h3>\n \n <p>Anatomical variations during radiotherapy fractions can lead to deviations in radiation delivery. Predicting these changes may benefit adaptive radiotherapy (ART) in nasopharyngeal cancer.</p>\n </section>\n \n <section>\n \n <h3> Purpose</h3>\n \n <p>This study proposes a vector quantized variational autoencoder (VQ-VAE) based generative model to predict anatomical changes in nasopharyngeal cancer patients.</p>\n </section>\n \n <section>\n \n <h3> Methods</h3>\n \n <p>The model integrates a VQ-VAE with Adaptive Instance Normalization (AdaIN). VQ-VAE encodes anatomical structures from planning CT images, while a convolutional neural network (CNN) extracts latent codes that capture potential anatomical variations. AdaIN then modulates the VQ-VAE's latent space to generate daily CT images that reflect these anatomical changes. The model was trained and validated on 522 CT images from 90 nasopharyngeal cancer patients and tested using 102 CT images from 18 patients. The quality of the generated images was evaluated through visual inspection, while the model's accuracy was assessed by comparing the predicted and actual volumes of the parotid and submandibular glands at both individual and population levels.</p>\n </section>\n \n <section>\n \n <h3> Results</h3>\n \n <p>For individual patients, Mann–Whitney and Kruskal–Wallis tests found no significant differences in organ-at-risk (OAR) volume distributions between generated and actual daily CT images. At the population level, predicted mean ROI volumes (parotid glands: 26.9 ± 2.1 cm<sup>3</sup>; submandibular glands: 7.0 ± 0.71 cm<sup>3</sup>) closely matched ground truth values (parotid: 29.5 ± 3.2 cm<sup>3</sup>; submandibular: 7.2 ± 0.67 cm<sup>3</sup>) and outperformed the previous Daily Anatomy Model (DAM) model (parotid: 20.4 ± 1.9 cm<sup>3</sup>; submandibular: 6.2 ± 0.6 cm<sup>3</sup>). The Pearson correlation coefficients between actual and generated daily CT ROI volumes were 0.92, 0.87, 0.89, and 0.93 for right parotid, left parotid, right submandibular, and left submandibular, respectively.</p>\n </section>\n \n <section>\n \n <h3> Conclusions</h3>\n \n <p>The VQ-VAE model effectively predicts anatomical changes during radiotherapy based on planning CT, demonstrating its potential to inform adaptive decision-making in radiotherapy.</p>\n </section>\n </div>","PeriodicalId":18384,"journal":{"name":"Medical physics","volume":"52 9","pages":""},"PeriodicalIF":3.2000,"publicationDate":"2025-09-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Medical physics","FirstCategoryId":"3","ListUrlMain":"https://aapm.onlinelibrary.wiley.com/doi/10.1002/mp.18120","RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"RADIOLOGY, NUCLEAR MEDICINE & MEDICAL IMAGING","Score":null,"Total":0}

引用次数: 0

Abstract

Background

Anatomical variations during radiotherapy fractions can lead to deviations in radiation delivery. Predicting these changes may benefit adaptive radiotherapy (ART) in nasopharyngeal cancer.

Purpose

This study proposes a vector quantized variational autoencoder (VQ-VAE) based generative model to predict anatomical changes in nasopharyngeal cancer patients.

Methods

The model integrates a VQ-VAE with Adaptive Instance Normalization (AdaIN). VQ-VAE encodes anatomical structures from planning CT images, while a convolutional neural network (CNN) extracts latent codes that capture potential anatomical variations. AdaIN then modulates the VQ-VAE's latent space to generate daily CT images that reflect these anatomical changes. The model was trained and validated on 522 CT images from 90 nasopharyngeal cancer patients and tested using 102 CT images from 18 patients. The quality of the generated images was evaluated through visual inspection, while the model's accuracy was assessed by comparing the predicted and actual volumes of the parotid and submandibular glands at both individual and population levels.

Results

For individual patients, Mann–Whitney and Kruskal–Wallis tests found no significant differences in organ-at-risk (OAR) volume distributions between generated and actual daily CT images. At the population level, predicted mean ROI volumes (parotid glands: 26.9 ± 2.1 cm³; submandibular glands: 7.0 ± 0.71 cm³) closely matched ground truth values (parotid: 29.5 ± 3.2 cm³; submandibular: 7.2 ± 0.67 cm³) and outperformed the previous Daily Anatomy Model (DAM) model (parotid: 20.4 ± 1.9 cm³; submandibular: 6.2 ± 0.6 cm³). The Pearson correlation coefficients between actual and generated daily CT ROI volumes were 0.92, 0.87, 0.89, and 0.93 for right parotid, left parotid, right submandibular, and left submandibular, respectively.

Conclusions

The VQ-VAE model effectively predicts anatomical changes during radiotherapy based on planning CT, demonstrating its potential to inform adaptive decision-making in radiotherapy.

Abstract Image

查看原文本刊更多论文

用矢量量化变分自编码器生成模型预测放射治疗中的解剖变异

背景：放射治疗过程中解剖结构的变化可能导致放射传递的偏差。预测这些变化可能有利于鼻咽癌的适应性放疗（ART）。目的提出一种基于矢量量化变分自编码器（VQ-VAE）的生成模型来预测鼻咽癌患者的解剖变化。方法将VQ-VAE与自适应实例归一化（AdaIN）相结合。VQ-VAE从规划CT图像中编码解剖结构，而卷积神经网络（CNN）提取捕捉潜在解剖变化的潜在代码。然后，AdaIN调整VQ-VAE的潜在空间，生成反映这些解剖变化的日常CT图像。对90例鼻咽癌患者的522张CT图像进行了训练和验证，并对18例鼻咽癌患者的102张CT图像进行了测试。通过目视检查来评估生成图像的质量，而通过比较个体和群体水平上腮腺和下颌下腺的预测和实际体积来评估模型的准确性。结果对于个别患者，Mann-Whitney和Kruskal-Wallis试验发现生成和实际日常CT图像之间的器官危险（OAR）体积分布无显著差异。在总体水平上，预测的平均ROI体积（腮腺：26.9±2.1 cm3；下颌骨腺：7.0±0.71 cm3）与基础真值（腮腺：29.5±3.2 cm3；下颌骨：7.2±0.67 cm3）非常匹配，优于之前的每日解剖模型（DAM）模型（腮腺：20.4±1.9 cm3；下颌骨：6.2±0.6 cm3）。右侧腮腺、左侧腮腺、右侧下颌骨和左侧下颌骨实际和生成的每日CT ROI体积之间的Pearson相关系数分别为0.92、0.87、0.89和0.93。结论VQ-VAE模型基于规划CT有效预测放疗过程中的解剖变化，显示了其为放疗适应性决策提供信息的潜力。

本文章由计算机程序翻译，如有差异，请以英文原文为准。

求助全文

约1分钟内获得全文求助全文

来源期刊

Medical physics 医学-核医学

CiteScore

6.80

自引率

15.80%

发文量

660

审稿时长

1.7 months

期刊介绍： Medical Physics publishes original, high impact physics, imaging science, and engineering research that advances patient diagnosis and therapy through contributions in 1) Basic science developments with high potential for clinical translation 2) Clinical applications of cutting edge engineering and physics innovations 3) Broadly applicable and innovative clinical physics developments Medical Physics is a journal of global scope and reach. By publishing in Medical Physics your research will reach an international, multidisciplinary audience including practicing medical physicists as well as physics- and engineering based translational scientists. We work closely with authors of promising articles to improve their quality.