文献速递泌尿生殖腹部CT

深度学习作为无创工具在CT上区分肌层浸润性与非肌层浸润性膀胱癌的应用

Application of deep learning as a noninvasive tool to differentiate muscle-invasive bladder cancer and non-muscle-invasive bladder cancer with CT.
PubMed 文献 · Eur J Radiol · 2021-03-19

本研究基于369例膀胱癌患者的1200幅增强CT横断面图像,构建深度学习卷积神经网络系统以区分肌层浸润性膀胱癌(MIBC)与非肌层浸润性膀胱癌(NMIBC)。小规模DL-CNN在验证集和测试集的AUROC分别为0.946和0.998,八种预训练模型在测试集的AUROC为0.762至0.997,其中VGG16模型AUROC最高(0.997),敏感度和特异度分别为0.889和0.989。结论认为该系统能有效分类NMIBC与MIBC,早期应用可能辅助病理检查提高诊断准确性。

要点

  • 增强CT图像可用于深度学习区分MIBC与NMIBC,测试集AUROC可达0.997。
  • VGG16预训练模型表现最佳,敏感度0.889、特异度0.989。
  • 该系统可能作为无创工具辅助病理检查,提高诊断准确性。

核心结论

  • 基于增强CT图像的DL-CNN系统能成功分类NMIBC与MIBC,AUROC良好。
  • 早期应用该系统可能辅助病理检查,提高诊断准确性。

研究设计与样本

  • 回顾性研究,纳入2015年1月至2018年6月接受根治性膀胱切除术的369例膀胱癌患者。
  • 共获取1200幅横断面CT图像,包括249例NMIBC序列和120例MIBC序列。
  • 图像按70%、15%、15%比例随机分为训练集、验证集和测试集。
  • 构建1个小规模DL-CNN(含4个卷积和最大池化层)及8个基于ImageNet预训练的DL-CNN。
  • 参考标准为根治性膀胱切除术后的病理结果。

诊断标准与关键数值

  • 小规模DL-CNN在验证集AUROC为0.946,测试集AUROC为0.998。
  • 8个预训练深度学习算法在测试集AUROC范围为0.762至0.997。
  • VGG16模型在8个算法中AUROC最大为0.997,敏感度0.889,特异度0.989。

临床应用与阅片要点

  • 增强CT图像可用于深度学习模型区分MIBC与NMIBC,辅助术前分期。
  • VGG16预训练模型表现出高特异度,可作为候选模型参考。
  • 模型输出可辅助病理检查,但需结合临床和病理结果综合判断。

局限性

  • 单中心回顾性研究,样本来自单一时段和单一机构。
  • 样本量有限(369例),需外部验证确认泛化能力。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

英文原文(摘要)· 点击展开

OBJECTIVE: To construct a deep-learning convolution neural network (DL-CNN) system for the differentiation of muscle-invasive bladder cancer (MIBC) and non-muscle-invasive bladder cancer (NMIBC) on contrast-enhanced computed tomography (CT) images in patients with bladder cancer.

MATERIALS AND METHODS: A total of 1200 cross-sectional CT images were obtained from 369 patients with bladder cancer receiving radical cystectomy from January 2015 to June 2018, including 249 non-muscle-invasive bladder cancer (NMIBC) series and 120 muscle-invasive bladder cancer (MIBC) series. All eligible images were distributed randomly into the training, validation, and testing cohorts with ratios of 70 %, 15 %, and 15 %, respectively. We developed one small DL-CNN containing four convolutional and max pooling layers and eight DL-CNNs with pretrained bases from the ImageNet dataset to differentiate NMIBC from MIBC. The intermediate activations were applied on the test dataset to visualize how successive DL-CNN layers transform their input.

RESULTS: The area under the receiver operating characteristic curve (AUROC) of the validation and testing datasets for the small DL-CNN was 0.946 and 0.998, respectively. The AUROCs of eight deep learning algorithms with pretrained bases ranged from 0.762 to 0.997 in the testing dataset. The VGG16 model had the largest AUROC of 0.997 among the eight algorithms with a sensitivity and specificity of 0.889 and 0.989. The independent features encoded by the small DL-CNN filters were displayed as assemblies of individual channels.

CONCLUSION: Based on contrast-enhanced CT images, our DL-CNN system could successfully classify NMIBC and MIBC with favorable AUROC in patients with bladder cancer. The application of our system in early stage might assist the pathological examination for the improvement of diagnostic accuracy.

来源:Eur J Radiol,查看原文。版权归原作者所有。

引用

  1. Yang Y, Zou X, Wang Y, et al. Application of deep learning as a noninvasive tool to differentiate muscle-invasive bladder cancer and non-muscle-invasive bladder cancer with CT. Eur J Radiol. 2021. PMID: 33798819. doi:10.1016/j.ejrad.2021.109666

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。