Department of Radiation Oncology and Winship Cancer Institute, Emory University, Atlanta, GA, United States of America.
Phys Med Biol. 2020 Oct 9;65(20):205003. doi: 10.1088/1361-6560/aba410.
Multi-needle localization in ultrasound (US) images is a crucial step of treatment planning for US-guided prostate brachytherapy. However, current computer-aided technologies are mostly focused on single-needle digitization, while manual digitization is labor intensive and time consuming. In this paper, we proposed a deep learning-based workflow for fast automatic multi-needle digitization, including needle shaft detection and needle tip detection. The major workflow is composed of two components: a large margin mask R-CNN model (LMMask R-CNN), which adopts the lager margin loss to reformulate Mask R-CNN for needle shaft localization, and a needle based density-based spatial clustering of application with noise algorithm which integrates priors to model a needle in an iteration for a needle shaft refinement and tip detections. Besides, we use the skipping connection in neural network architecture to improve the supervision in hidden layers. Our workflow was evaluated on 23 patients who underwent US-guided high-dose-rate (HDR) prostrate brachytherapy with 339 needles being tested in total. Our method detected 98% of the needles with 0.091 ± 0.043 mm shaft error and 0.330 ± 0.363 mm tip error. Compared with only using Mask R-CNN and only using LMMask R-CNN, the proposed method gains a significant improvement on both shaft error and tip error. The proposed method automatically digitizes needles per patient with in a second. It streamlines the workflow of transrectal ultrasound-guided HDR prostate brachytherapy and paves the way for the development of real-time treatment planning system that is expected to further elevate the quality and outcome of HDR prostate brachytherapy.
多针在超声（US）图像中的定位是 US 引导前列腺近距离放射治疗计划治疗的关键步骤。然而，当前的计算机辅助技术主要集中在单针数字化上，而手动数字化既费力又耗时。在本文中，我们提出了一种基于深度学习的快速自动多针数字化工作流程，包括针轴检测和针尖检测。主要工作流程由两部分组成：一个大边缘掩模 R-CNN 模型（LMMask R-CNN），它采用更大的边缘损失来重新定义掩模 R-CNN 以进行针轴定位，以及一个基于针的基于密度的空间聚类应用噪声算法，该算法集成了先验知识来迭代建模针以细化针轴并进行针尖检测。此外，我们在神经网络架构中使用跳过连接来改进隐藏层的监督。我们的工作流程在 23 名接受 US 引导高剂量率（HDR）前列腺近距离放射治疗的患者中进行了评估，总共测试了 339 根针。我们的方法检测到 98%的针，其轴误差为 0.091 ± 0.043 毫米，尖端误差为 0.330 ± 0.363 毫米。与仅使用掩模 R-CNN 和仅使用 LMMask R-CNN 相比，所提出的方法在轴误差和尖端误差方面都有显著提高。所提出的方法可在一秒钟内为每个患者自动数字化针。它简化了经直肠超声引导 HDR 前列腺近距离放射治疗的工作流程，为开发实时治疗计划系统铺平了道路，预计这将进一步提高 HDR 前列腺近距离放射治疗的质量和效果。
