Ishiwata Tsukasa, Inage Terunaga, Aragaki Masato, Gregor Alexander, Chen Zhenchian, Bernards Nicholas, Kafi Kamran, Yasufuku Kazuhiro
Division of Thoracic Surgery, Toronto General Hospital, University Health Network, Toronto, Ontario, Canada.
Imagia Cybernetics, Montreal, Québec, Canada.
JTCVS Tech. 2024 Sep 19;28:151-161. doi: 10.1016/j.xjtc.2024.09.008. eCollection 2024 Dec.
Endobronchial ultrasound-guided transbronchial needle aspiration is a vital tool for mediastinal and hilar lymph node staging in patients with lung cancer. Despite its high diagnostic performance and safety, it has a limited negative predictive value. Our objective was to evaluate the diagnostic performance of deep learning-based prediction of lung cancer lymph node metastases using convolutional neural networks developed from automatically extracted images of endobronchial ultrasound videos without supervision of the lymph node location.
Patient and lymph node data were collected from a single-center database. The diagnosis of metastasis was confirmed with endobronchial ultrasound-guided transbronchial needle aspiration and/or surgically resected specimens; the diagnosis of normal lymph node was confirmed with surgically resected specimens only. An annotation system facilitated automated image extraction from endobronchial ultrasound videos. Image frames were randomly selected and split into training and validation datasets on a per-patient basis. A deep learning model with convolutional neural networks, SqueezeNet, was used for image classification via transfer learning based on pretraining from ImageNet. Adaptive moment estimation and stochastic gradient descent were applied as optimizers.
SqueezeNet, with adaptive moment estimation, achieved a sensitivity, specificity, accuracy, positive predictive value, and negative predictive value of 96.7% each after 300 epochs, whereas SqueezeNet with stochastic gradient descent achieved 91.1% each. However, SqueezeNet with stochastic gradient descent demonstrated more stable performance than with adaptive moment estimation.
Deep learning-based image classification using convolutional neural networks showed promising diagnostic accuracy for lung cancer nodal metastasis. Future clinical trials are warranted to validate the algorithm's efficacy in a prospective, large-cohort study.
支气管内超声引导下经支气管针吸活检术是肺癌患者纵隔和肺门淋巴结分期的重要工具。尽管其具有较高的诊断性能和安全性，但其阴性预测值有限。我们的目的是评估基于深度学习的肺癌淋巴结转移预测的诊断性能，该预测使用从支气管内超声视频自动提取的图像开发的卷积神经网络，且无需淋巴结位置的监督。
从单中心数据库收集患者和淋巴结数据。通过支气管内超声引导下经支气管针吸活检术和/或手术切除标本确认转移诊断；仅通过手术切除标本确认正常淋巴结诊断。一个注释系统有助于从支气管内超声视频中自动提取图像。图像帧在每个患者的基础上随机选择并分为训练和验证数据集。使用具有卷积神经网络的深度学习模型SqueezeNet，通过基于ImageNet预训练的迁移学习进行图像分类。应用自适应矩估计和随机梯度下降作为优化器。
采用自适应矩估计的SqueezeNet在300个轮次后，灵敏度、特异度、准确率、阳性预测值和阴性预测值均达到96.7%，而采用随机梯度下降的SqueezeNet各项指标均为91.1%。然而，采用随机梯度下降的SqueezeNet比采用自适应矩估计的表现出更稳定的性能。
使用卷积神经网络的基于深度学习的图像分类对肺癌淋巴结转移显示出有前景的诊断准确性。未来有必要进行临床试验，以前瞻性、大样本队列研究验证该算法的疗效。
