文献速递：深度学习胰腺癌诊断--胰腺癌在CT扫描中通过深度学习检测：一项全国性的基于人群的研究
Title
题目
Pancreatic Cancer Detection on CT Scans with Deep
Learning: A Nationwide Population-based Study
胰腺癌在CT扫描中通过深度学习检测：一项全国性的基于人群的研究
01
文献速递介绍
在所有癌症类型中胰腺癌（PC）具有最差的五年生存率预计到2030年该病症将占美国癌症死亡原因排名第二的位置
近期研究表明，在医学图像分析领域中深度学习（DL）技术展现出巨大潜力
Background
背景
About one in five pancreatic tumors under 2 centimeters are missed during abdominal CT scans.
大约40%的小于2厘米的胰腺肿瘤在腹部CT扫描中被遗漏。
Conclusions
结论
A deep-learning-enabled diagnostic system has been developed to accurately identify pancreatic cancer from CT imaging procedures, achieving a sensitivity level that is deemed acceptable for detecting tumors below 2 cm.
深度学习驱动的解决方案在CT影像分析中实现了精准诊断，并对直径小于2厘米的病变区域具有较高的检出能力。
Results
结果
A total of 546 patients with pancreatic cancer (mean age, 65 years ± 12 [SD], 297 men) and 733 control subjects were ran domly divided into training, validation, and test sets. In the internal test set, the DL tool achieved 89.9% (98 of 109; 95% CI: 82.7, 94.9) sensitivity and 95.9% (141 of 147; 95% CI: 91.3, 98.5) specificity (area under the receiver operating characteristic curve [AUC], 0.96; 95% CI: 0.94, 0.99), without a significant difference (P = .11) in sensitivity compared with the original radiologist report (96.1% [98 of 102]; 95% CI: 90.3, 98.9). In a test set of 1473 real-world CT studies (669 malignant, 804 control) from institutions throughout Taiwan, the DL tool distinguished between CT malignant and control studies with 89.7% (600 of 669; 95% CI: 87.1, 91.9) sensitivity and 92.8% specificity (746 of 804; 95% CI: 90.8, 94.5) (AUC, 0.95; 95% CI: 0.94, 0.96), with 74.7% (68 of 91; 95% CI: 64.5, 83.3) sensitivity for malignancies smaller than 2 cm.
共有546例胰腺癌病例（平均年龄为65岁±约12年标准差，在此群体中有约三分之二是男性）按照随机分配的方式被划分为训练组、验证组及测试组
Method
方法
retrospectively gathered contrast-enhanced CT scans targeting patients with pancreatic cancer diagnosed between January 2006 and July 2018 were systematically evaluated against CT imaging findings from individuals exhibiting a normal pancreatic anatomy (control group), which were collected between January 2004 and December 2019. An integrated system incorporating both a segmentation convolutional neural network (CNN) for image analysis and an ensemble of five CNN classifiers for diagnostic support was constructed, then rigorously tested within an internal validation cohort as well as across a nationwide real-world testing platform. A McNemar statistical test was employed to assess whether the sensitivities of CAD-guided diagnostics differ significantly from those achieved by human radiologists interpreting CT images.
回顾性采集了2006年1月至2018年7月期间被诊断为胰腺癌患者的增强对比CT影像资料，并将其与2004年1月至2019年12月间经诊断患有正常胰腺个体（对照组）的标准CT检查进行比较。研制并评估了一个完整的端到端系统架构, 包含一个分割式卷积神经网络模型以及集成五个不同模型的分类器, 在内部测试集以及全国范围内的真实案例验证中均表现优异。通过McNemar统计检验方法评估了计算机辅助检测系统与放射科医生解读结果在灵敏度方面的差异
Figure
图
The dataset figure 1 details the data collection efforts for the (A) segmentation technique and the (B) local and national datasets used in classification approaches.
图1：(A) 分割模型的数据集以及 (B) 分类模型的本地和全国范围数据集。
Chart 2 illustrates the workflow of a computer-aided detection tool utilizing deep learning technology. The segmentation masks generated by a segmentation convolutional neural network were transmitted to classification CNNs, incorporating both the pancreas and tumor (if present), without distinguishing between them. Solid arrows signify that a computer-aided detection system produces outputs based on these segmented masks.
图2展示了基于深度学习技术实现的计算机辅助检测系统工作流程图。该系统中分割卷积神经网络（CNN）将分割掩模传递至分类级的过程中包含了胰腺及可能存在的肿瘤区域。值得注意的是该系统未对胰腺与肿瘤间的细微差别进行单独识别处理
Figure 3 illustrates the Receiver operating characteristic (ROC) curves for the classification models across three distinct testing datasets: (A) training and validation sets within the dataset, (B) local testing datasets, and (C), which covers nationwide testing. The abbreviation CNN stands for convolutional neural network as shown in Fig 3, which continues.
图3展示了分类模型的receiver operating characteristic curve，在以下三个数据集中分别展示：(A) 训练与验证集、(B)本地测试集以及(C)全国范围测试集中。其中CNN代表卷积神经网络（图3继续）。
Figure 3 continued: Selected CT scans arranged in a left column showing tumors located at the pancreas' head, body, and tail correspond to manual segmentation results performed by radiologists in the middle column and model predictions displayed in the right column. Blue outlines denote the pancreas while yellow outlines indicate tumors.
图3 (续)：典型的CT扫描（左栏）展示了胰腺（D）头部、（E）体部及（F）尾部的肿瘤区域。对比显示，在肿瘤位置上，放射科医师的手工分割（中栏）与分割模型预测的结果（右栏）存在对应关系。其中，蓝色轮廓代表胰腺；而黄色轮廓代表肿瘤。
_Figure 4: 图4展示了分割模型中出现的误报（A、B）和假阳性（C、D）肿瘤分割情况。蓝色和黄色轮廓分别指示正常胰腺以及通过分割模型实现的肿瘤轮廓。左侧列中的图片均为未经注释的人工断层扫描CT影像。（A、B）图中未被分割的肿瘤（红轮廓）显示胰腺癌的相关迹象：包括胰腺导管扩张并突然截断（A图箭头所示）、以及胰腺实质性萎缩并伴随胰腺导管扩张（B图箭头所示）。C图中由于肝静脉球形虫病继发性 portal 罗所以我们错误地将它们分割为肿瘤。（D）图显示，在接受胆道支架置入以解除来自肝细胞癌阻塞的黄疸过程中所见胰腺实质与支架相邻区域被错误地划分为肿瘤区域。
图4展示了分割系统的漏分情况(A, B)与误分情况(C, D)在胰腺癌分割中的应用效果。(A, B)中未被分割的卷积神经网络对胰腺癌表现出了较高的诊断能力：其将胰管扩张及突然截止（A图中标注）识别为肿瘤特征；而在胰腺实质萎缩与胰管扩张（B图中标注）处同样表现出敏感性。(C)特发性门静脉血栓所形成的侧支静脉被误判为肿瘤区域。(D)为了缓解肝细胞癌导致的梗阻性黄疸，在手术切口处放置的胆道支架（箭头指示位置）附近出现的胰腺实质同样被误认为是肿瘤区域
Figure 5: 对胰腺非肿瘤部分的研究与胰腺癌前表现的关系通过分类模型进行分析 蓝线表示通过分类模型研究的胰腺部位 肿瘤未被分割模型识别到（未标注区域）
图5
Table
表
Table 1: 胰腺癌研究对象特征与对照组在本地数据集中的特点
表1：本地数据集中胰腺癌和对照组的特征
Table 2.1: Positive Predictive Value (PPV) Based on Types of Classification Using Convolutional Neural Networks Diagnosed with Pancratic Cancer
表2：根据预测为胰腺癌的分类卷积神经网络的数量的阳性似然比
Table 3: Computer-Aided Detection Tool's Performance and Radiologists' Results Derived from the Original Report When Distinguishing CT Scans with or without Pancreatic Cancer]
表3：源自原始数据来源，在有无胰腺癌的CT研究中进行鉴别比较的是放射科医生与计算机辅助检测技术的表现。
Table 4: The Sensitivity of the Computer-Assisted Detection Tool in the Context of Radiologists Categorized According to Tumor Progression Stages and Sizes
表4：基于肿瘤分期与大小分层的计算机辅助诊断工具与放射科医师敏感性