基于剂量预测和自动勾画技术的PET/CT器官内照射剂量率快速评估方法

Fast estimation of internal irradiation dose rate in PET/CT imaging using deep learning-based dose prediction combined with auto-segmentation technique

目的:实现一种基于深度学习的剂量预测和自动勾画技术的正电子发射断层成像(PET)/CT检查器官内照射剂量率的快速评估方法。方法:首先基于患者特定时刻的PET/CT图像,使用蒙特卡罗程序GATE进行内照射剂量率计算,获得每个患者的剂量率分布图。随后,基于U-Net构建深度神经网络,将患者的CT和PET图像作为输入,GATE计算的剂量率图作为金标准进行训练。训练后的深度学习模型能够根据患者的CT和PET图像预测对应的剂量率分布。同时,使用勾画软件DeepViewer对患者CT图像中的器官和组织进行自动勾画,结合预测得到的剂量率分布结果计算相应器官和组织的吸收剂量率。使用50名患者的PET/CT数据,其中10份用于测试,其余40份进行4折交叉训练,每次使用30份用于训练,10份用于验证。将测试集结果与GATE和GPU蒙特卡罗工具ARCHER-NM进行对比。结果:在自动勾画软件DeepViewer勾画的24个器官中,绝大部分器官的深度学习预测剂量率与GATE计算结果偏差在±10%以内。其中大脑、心脏、肝脏、左肺、右肺的平均偏差分别为3.3%、1.1%、1.0%、-1.1%、0.0%,与GATE具有较好的一致性。使用GATE程序进行每名患者的内照射剂量率计算平均用时8.91 h,而使用深度神经网络模型进行内照射剂量率预测平均每名患者用时15.1 s,平均加速比达到2 120倍。和ARCHER-NM的对比表明,基于深度学习方法的剂量率预测具有速度优势,但在结果的可解释性方面还需要改善。结论:利用深度学习预测和自动勾画技术可以从PET/CT图像快速得到剂量率分布,有望作为一种PET/CT内照射剂量率快速评估方法,为临床核医学快速、实时地计算人体内照射吸收剂量提供一种新的解决方案。

Objective To realize the fast estimation of internal irradiation dose rate in PET/CT imaging using the combination of deep learning-based dose prediction and auto-segmentation technique. Methods Based on the PET/CT images of the patients at a specific moment, Monte Carlo simulation software GATE was used to calculate the internal irradiation dose rate,and obtain the dose rate distribution map of each patient. The PET and CT image patches were used as inputs for the training of a deep neural network constructed based on U-Net network, while internal irradiation dose rate map calculated by Monte Carlo simulation software GATE was given as ground truth. The trained deep learning model could predict the dose rate map according to the PET/CT images. Meanwhile, the radiosensitive organs and tissues in the CT images were automatically segmented using DeepViewer. The absorbed dose rates of the corresponding organs and tissues were calculated based on organ segmentation results and the predicted dose rate distribution. The PET/CT images of 50 patients were used in the study.Ten of which were used as testing set, and the others were used for 4-fold cross-validation training, with 30 for training and10 for validation in each fold. The predicted results were compared with the results obtained by GATE and ARCHER-NM(a GPU-accelerated Monte Carlo dose calculation module). Results For most of the 24 organs segmented by DeepViewer, the relative differences between predicted dose rate and GATE simulation results were within ±10%. Specifically, the average relative differences of brain, heart, liver, left and right lungs were 3.3%, 1.1%, 1.0%,-1.1% and 0.0%, respectively,indicating a good consistency between dose rate prediction and GATE simulation. For each patient, the deep learning-based prediction costs 15.1 s on average for the estimation of internal irradiation dose rate, while the GATE simulation costs 8.91 h.The calculation speed was increased by a factor of 2 120. The comparison between deep learning-based prediction and ARCHER-NM showed that the deep learning-based prediction had an advantage of execution time, while its interpretability needed further improvement. Conclusion The combination of deep learning-based dose prediction and auto-segmentation technique is expected to be a method for the rapid estimation of internal irradiation dose rate in PET/CT imaging, and provide a solution to calculate the real-time internal absorbed dose rapidly for the practices of clinical nuclear medicine.