Department of Radiation Oncology, Josephine Ford Cancer Institute, Henry Ford Health System, Detroit, Michigan.
Department of Radiation Oncology, Josephine Ford Cancer Institute, Henry Ford Health System, Detroit, Michigan.
Int J Radiat Oncol Biol Phys. 2019 Jul 15;104(4):924-932. doi: 10.1016/j.ijrobp.2019.03.017. Epub 2019 Mar 16.
Recent advances in deep neural networks (DNNs) have unlocked opportunities for their application for automatic image segmentation. We have evaluated a DNN-based algorithm for automatic segmentation of the prostate gland on a large cohort of patient images.
Planning-CT data sets for 1114 patients with prostate cancer were retrospectively selected and divided into 2 groups. Group A contained 1104 data sets, with 1 physician-generated prostate gland contour for each data set. Among these image sets, 771 were used for training, 193 for validation, and 140 for testing. Group B contained 10 data sets, each including prostate contours delineated by 5 independent physicians and a consensus contour generated using the STAPLE method in the CERR software package. All images were resampled to a spatial resolution of 1 × 1 × 1.5 mm. A region (128 × 128 × 64 voxels) containing the prostate was selected to train a DNN. The best-performing model on the validation data sets was used to segment the prostate on all testing images. Results were compared between DNN and physician-generated contours using the Dice similarity coefficient, Hausdorff distances, regional contour distances, and center-of-mass distances.
The mean Dice similarity coefficients between DNN-based prostate segmentation and physician-generated contours for test data in Group A, Group B, and Group B-consensus were 0.85 ± 0.06 (range, 0.65-0.93), 0.85 ± 0.04 (range, 0.80-0.91), and 0.88 ± 0.03 (range, 0.82-0.92), respectively. The Hausdorff distance was 7.0 ± 3.5 mm, 7.3 ± 2.0 mm, and 6.3 ± 2.0 mm for Group A, Group B, and Group B-consensus, respectively. The mean center-of-mass distances for all 3 data set groups were within 5 mm.
A DNN-based algorithm was used to automatically segment the prostate for a large cohort of patients with prostate cancer. DNN-based prostate segmentations were compared to the consensus contour for a smaller group of patients; the agreement between DNN segmentations and consensus contour was similar to the agreement reported in a previous study. Clinical use of DNNs is promising, but further investigation is warranted.
深度学习神经网络（DNN）的最新进展为其在自动图像分割中的应用提供了机会。我们评估了一种基于 DNN 的算法，用于对大量患者图像的前列腺进行自动分割。
回顾性选择了 1114 例前列腺癌患者的计划 CT 数据集，并将其分为 2 组。A 组包含 1104 个数据集，每个数据集都有 1 个医生生成的前列腺轮廓。在这些图像集中，771 个用于训练，193 个用于验证，140 个用于测试。B 组包含 10 个数据集，每个数据集都包含由 5 位独立医生勾画的前列腺轮廓和使用 CERR 软件包中的 STAPLE 方法生成的共识轮廓。所有图像均重采样至空间分辨率为 1×1×1.5mm。选择包含前列腺的区域（128×128×64 个体素）来训练 DNN。在验证数据集上表现最佳的模型用于分割所有测试图像的前列腺。使用 Dice 相似系数、Hausdorff 距离、区域轮廓距离和质心距离比较 DNN 和医生生成轮廓之间的结果。
A 组、B 组和 B 组共识测试数据的基于 DNN 的前列腺分割与医生生成轮廓之间的平均 Dice 相似系数分别为 0.85±0.06（范围，0.65-0.93）、0.85±0.04（范围，0.80-0.91）和 0.88±0.03（范围，0.82-0.92）。A 组、B 组和 B 组共识的 Hausdorff 距离分别为 7.0±3.5mm、7.3±2.0mm 和 6.3±2.0mm。所有 3 个数据集组的质心距离平均值均在 5mm 以内。
使用基于 DNN 的算法自动分割了大量前列腺癌患者的前列腺。将基于 DNN 的前列腺分割与较小患者组的共识轮廓进行了比较；DNN 分割与共识轮廓之间的一致性与之前的研究报告中的一致性相似。DNN 的临床应用很有前景，但需要进一步研究。
