【摘要】 目的：建立一种深度学习去噪模型，加速电子射野影像装置（EPID）透射剂量图像的蒙特卡罗模拟。方法：收集5野调强放疗的肺癌患者100例，共500个EPID射野，随机选取400个射野作为训练集，50个射野作为验证集，剩下的50个射野作为测试集。使用GPU加速的蒙特卡罗剂量计算引擎ARCHER模拟低粒子数（1×10~7）和高粒子数（1×10~9）的EPID透射剂量图像数据集。开发基于Swin Transformer和U-Net的去噪网络模型（SUNet），分别将低粒子数图像和高粒子数图像作为输入和输出对网络进行训练，并用该模型对测试集的低粒子数EPID图像进行去噪。利用结构相似性系数（SSIM）、峰值信噪比（PSNR）和γ通过率（3%/2 mm）等参数评估去噪效果，分析结合SUNet模型的蒙特卡罗模拟的计算效率。结果：SUNet去噪后图像与原始低粒子数图像相比，图像的质量大幅改善、噪点更少、剂量分布更平滑。SUNet去噪后图像与高粒子数图像相比，平均SSIM大于0.9，平均PSNR大于32 dB，平均γ通过率超过90%。结合SUNet的蒙特卡罗方法模拟一个EPID透射剂量图像仅需1.88 s，与高粒子数蒙特卡罗模拟相比计算效率提高大约40倍。结论：基于深度学习的去噪模型在保持图像质量和剂量准确性的同时大幅提高EPID透射剂量图像的蒙特卡罗模拟速度，为基于EPID的在体剂量验证提供可能。更多还原

【Abstract】 Objective To develop a deep learning-based denoising model for accelerating Monte Carlo(MC) simulation of electronic portal imaging device(EPID) transit dose images. Methods A total of 500 EPID fields were collected from 100 lung cancer patients undergoing 5-field intensity-modulated radiotherapy, with 400 fields randomly selected as training set,50 fields as validation set, and 50 fields as test set. EPID transit dose image datasets with low particle counts(1×10~7) and high particle counts(1×10~9) were simulated using the GPU-accelerated MC dose calculation engine ARCHER. A denoising network model named SUNet was constructed based on Swin Transformer and U-Net, and trained using low-particle-count images as input and high-particle-count images as output. Following training, SUNet model was used to denoise low-particlecount EPID images in the test set. Denoising performance was evaluated using structural similarity index(SSIM), peak signal-to-noise ratio(PSNR), and Gamma passing rates(3%/2 mm), and the computational efficiency of MC simulation combined with SUNet model was analyzed. Results Compared with the original low-particle-count images, the SUNetdenoised images showed significantly improved quality, reduced noise points, and smoother dose distribution. When benchmarked against high-particle-count images, the SUNet-denoised images achieved an average SSIM greater than 0.9, an average PSNR higher than 32 dB, and an average gamma passing rate exceeding 90%. The MC simulation combined with SUNet model required only 1.88 s to simulate a single EPID transit dose image, representing an approximate 40-fold improvement in computational efficiency as compared with high-particle-count MC simulation. Conclusion The deep learning-based denoising model substantially accelerates MC simulation of EPID transit dose images while preserving both image quality and dose accuracy, which provides possibilities for EPID-based in vivo dose verification.更多还原