兰州财经大学机构知识库

肺癌是世界第二大常见癌症，且在所有癌症中致死率最高。全身肿瘤体积（whole-body metabolic tumor volume, MTVwb）是用于预测肺癌患者生存时间的独立变量，目前该变量的计算依赖医生手动勾画的肿瘤标签。勾画肿瘤标签是一个费时费力、重复性差的工作，而且医生在勾画时，由于肿瘤边界的模糊性，会产生主观差异。使用计算机自动分割肿瘤能辅助医生对癌症的诊断。现有PET/CT肺癌肿瘤分割方法在胸部图像表现较好，但没有考虑肿瘤向全身转移的情况。由于全身图像远比胸部图像复杂，直接将这些方法套用在全身图像上并不能取得很好的效果，因此需要针对全身PET/CT的肿瘤分割方法。

本文提出了一种位置监督的多分支神经网络（position supervised multi-branch network, PSMB-Net）用于从全身PET/CT中分割肿瘤。方法从三个角度进行研究：第一，CT图像包含了人体固有的结构规律，CT切片稳定的相对位置体现了这种规律，对该相对位置的学习可以辅助PET/CT肿瘤分割；第二，全身PET图像中存在与肿瘤高度相似的正常代谢活跃组织，胸外CT中肿瘤与周围正常组织的边界极难划分，使得全身图像的规律远比胸部图像复杂。通过观察发现全身PET/CT可以以肺部下界为标准，分为上下两个半身，使得每个半身图像内部的规律比全身图像更为简单。基于此，本文定义了两个基于位置的解码器，分别学习人体上下两个半身的特征，用两个门信号控制它们的优化过程，门信号将切片的相对位置转换为切片在输出中的权重；第三，全身PET/CT图像大肿瘤与小肿瘤的像素个数差距悬殊，使得在标准的交叉熵分割损失的优化下，模型对小肿瘤的预测不够全面。在标准交叉熵中加入一个肿瘤级的平衡系数可以缓解这个问题。

实验使用480位肺癌患者的全身PET/CT测试该方法的性能。在5折交叉测试中，分别计算图像分割指标和生存分析指标。本文提出的PSMB-Net在Dice、Sensitivity、Precision上分别达到了0.580、0.616和0.688，均超过了所比较的方法。在生存分析中， PSMB-Net预测的MTVwb 计算出的C-index比来自医生勾画的MTVwb 在单变量下的Cox和RSF模型上仅低0.023和0.008。这些结果说明本文所提方法在全身PET/CT肿瘤分割任务上有一定价值。

Lung cancer is on the second rank of most common form of internal malignancies, and it is the leading cause of cancer related death in 2020. Whole-body metabolic tumor volume (MTVwb) is an independent prognostic factor for overall survival in lung cancer patients. Current calculation of this factor is depended on the tumor label provided by radiologists, which is labor consuming with poor reproducibility, and objective difference is introduced during the process of label delineation due to the blurry border. Automatic segmentation of lung cancer tumors to calculate MTVwb can assist radiologists in their assessment of lung cancer. Published methods appeared promising in segmentation accuracy in thoracic PET/CT, their methods do not quantify distant metastatic disease outside of the thoracic region. A simple application of these methods on whole-body PET/CT is not sufficient to generate acceptable results due to the greatly increased complexity, thus it is essential to search of a method for whole-body PET/CT segmentation.

We present a position supervised multi-branch neural network (PSMB-Net) to segment tumor from whole-body PET/CT images. We make our research from three points of view. First, CT images contain inherent structural pattern, and the relative position of CT slices is a form of this pattern. The learning of the relative position can assist tumor segmentation of PET/CT images. Second, normal organs and tissues can be metabolically active, and therefore appear similar to tumors on whole-body PET images, and the border of extra-thoracic tumors are difficult to identify on CT images, making the complexity of whole-body PET/CT is much higher than thoracic image. To accommodate the remarkable visual differences of the upper and lower semi-body, which is determined by the lower bound of lung, our design includes two decoders to learn features from the upper and lower semi-body separately. Two position-based gate signals transfer each position value to the weight of the corresponding slice in the output. Third, the voxel number of greater tumor versus smaller tumor are in greatly imbalanced, making the model biased to ignore smaller tumors in test stage, under the optimization of standard cross-entropy segmentation loss. We add a balance factor to cross-entropy loss function to alleviate this imbalance.

We test the performance of our method on a PET/CT image dataset including 480 lung cancer patients with 5-fold cross-validation using the metrics for image segmentation and survival analysis. Our PSMB-Net archived 0.580, 0.616, and 0.688, on Dice, Sensitivity and Precision, respectively, which archives higher performance than some state-of-art methods. The concordance indices (C-indices) for the univariate MTVwb segmented by PSMB-Net are only 0.023 and 0.008 lower than that calculated from tumor label, in Cox and RSF model, respectively. These results demonstrate that our method is of value for whole-body tumor segmentation of PET/CT images.