Jackson Price, Hardcastle Nicholas, Dawe Noel, Kron Tomas, Hofman Michael S, Hicks Rodney J
Sir Peter MacCallum Department of Oncology, The University of Melbourne, Melbourne, VIC, Australia.
Department of Molecular Imaging, Peter MacCallum Cancer Centre, Melbourne, VIC, Australia.
Front Oncol. 2018 Jun 14;8:215. doi: 10.3389/fonc.2018.00215. eCollection 2018.
Convolutional neural networks (CNNs) have been shown to be powerful tools to assist with object detection and-like a human observer-may be trained based on a relatively small cohort of reference subjects. Rapid, accurate organ recognition in medical imaging permits a variety of new quantitative diagnostic techniques. In the case of therapy with targeted radionuclides, it may permit comprehensive radiation dose analysis in a manner that would often be prohibitively time-consuming using conventional methods.
An automated image segmentation tool was developed based on three-dimensional CNNs to detect right and left kidney contours on non-contrast CT images. Model was trained based on 89 manually contoured cases and tested on a cohort of patients receiving therapy with Lu-prostate-specific membrane antigen-617 for metastatic prostate cancer. Automatically generated contours were compared with those drawn by an expert and assessed for similarity based on dice score, mean distance-to-agreement, and total segmented volume. Further, the contours were applied to voxel dose maps computed from post-treatment quantitative SPECT imaging to estimate renal radiation dose from therapy.
Neural network segmentation was able to identify right and left kidneys in all patients with a high degree of accuracy. The system was integrated into the hospital image database, returning contours for a selected study in approximately 90 s. Mean dice score was 0.91 and 0.86 for right and left kidneys, respectively. Poor performance was observed in three patients with cystic kidneys of which only few were included in the training data. No significant difference in mean radiation absorbed dose was observed between the manual and automated algorithms.
Automated contouring using CNNs shows promise in providing quantitative assessment of functional SPECT and possibly PET images; in this case demonstrating comparable accuracy for radiation dose interpretation in unsealed source therapy relative to a human observer.
卷积神经网络（CNN）已被证明是辅助目标检测的强大工具，并且——如同人类观察者一样——可以基于相对较少的参考对象队列进行训练。医学成像中快速、准确的器官识别可实现多种新的定量诊断技术。在使用靶向放射性核素进行治疗的情况下，它可以以一种使用传统方法通常会极其耗时的方式进行全面的辐射剂量分析。
基于三维CNN开发了一种自动图像分割工具，以检测非增强CT图像上的左右肾轮廓。该模型基于89个手动勾勒轮廓的病例进行训练，并在一组接受镥-前列腺特异性膜抗原-617治疗转移性前列腺癌的患者中进行测试。将自动生成的轮廓与专家绘制的轮廓进行比较，并根据骰子系数、平均一致距离和总分割体积评估其相似性。此外，将这些轮廓应用于从治疗后定量SPECT成像计算得到的体素剂量图，以估计治疗产生的肾脏辐射剂量。
神经网络分割能够在所有患者中高度准确地识别左右肾。该系统已集成到医院图像数据库中，在大约90秒内为选定的研究返回轮廓。左右肾的平均骰子系数分别为0.91和0.86。在3例多囊肾患者中观察到性能不佳，其中只有少数患者被纳入训练数据。手动算法和自动算法之间在平均辐射吸收剂量方面未观察到显著差异。
使用CNN进行自动轮廓勾勒在提供功能性SPECT以及可能的PET图像的定量评估方面显示出前景；在这种情况下，相对于人类观察者，在非密封源治疗中辐射剂量解释方面显示出相当的准确性。
