Department of Circulation and Medical Imaging, NTNU, Trondheim, Norway.
Department of Radiology and Nuclear Medicine, St. Olavs Hospital, Trondheim, Norway.
Int J Comput Assist Radiol Surg. 2019 Jun;14(6):977-986. doi: 10.1007/s11548-019-01948-8. Epub 2019 Mar 19.
Accurate lung cancer diagnosis is crucial to select the best course of action for treating the patient. From a simple chest CT volume, it is necessary to identify whether the cancer has spread to nearby lymph nodes or not. It is equally important to know precisely where each malignant lymph node is with respect to the surrounding anatomical structures and the airways. In this paper, we introduce a new data-set containing annotations of fifteen different anatomical structures in the mediastinal area, including lymph nodes of varying sizes. We present a 2D pipeline for semantic segmentation and instance detection of anatomical structures and potentially malignant lymph nodes in the mediastinal area.
We propose a 2D pipeline combining the strengths of U-Net for pixel-wise segmentation using a loss function dealing with data imbalance and Mask R-CNN providing instance detection and improved pixel-wise segmentation within bounding boxes. A final stage performs pixel-wise labels refinement and 3D instance detection using a tracking approach along the slicing dimension. Detected instances are represented by a 3D pixel-wise mask, bounding volume, and centroid position.
We validated our approach following a fivefold cross-validation over our new data-set of fifteen lung cancer patients. For the semantic segmentation task, we reach an average Dice score of 76% over all fifteen anatomical structures. For the lymph node instance detection task, we reach 75% recall for 9 false positives per patient, with an average centroid position estimation error of 3 mm in each dimension.
Fusing 2D networks' results increases pixel-wise segmentation results while enabling good instance detection. Better leveraging of the 3D information and station mapping for the detected lymph nodes are the next steps.
准确诊断肺癌对于选择治疗患者的最佳方案至关重要。从简单的胸部 CT 容积中，有必要确定癌症是否已扩散到附近的淋巴结。同样重要的是要准确了解每个恶性淋巴结相对于周围解剖结构和气道的位置。在本文中，我们介绍了一个包含十五个纵隔区域不同解剖结构注释的新数据集，包括大小不一的淋巴结。我们提出了一种 2D 流水线，用于对纵隔区域的解剖结构和潜在恶性淋巴结进行语义分割和实例检测。
我们提出了一种 2D 流水线，结合了 U-Net 在使用处理数据不平衡的损失函数进行像素级分割的优势和 Mask R-CNN 提供的实例检测和在边界框内改进的像素级分割。最后一个阶段使用沿切片维度的跟踪方法执行像素级标签细化和 3D 实例检测。检测到的实例由 3D 像素级掩模、边界体积和质心位置表示。
我们在十五例肺癌患者的新数据集上进行了五重交叉验证，验证了我们的方法。对于语义分割任务，我们在所有十五个解剖结构上达到了平均 Dice 得分为 76%。对于淋巴结实例检测任务，我们达到了每个患者 9 个假阳性的 75%召回率，每个维度的平均质心位置估计误差为 3 毫米。
融合 2D 网络的结果增加了像素级分割结果，同时实现了良好的实例检测。下一步是更好地利用 3D 信息和检测到的淋巴结的站映射。
